[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"portal-settings:stajic:sr":3,"public-menus:all":38,"post:agentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act:sr":205,"related:post:agentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act:sr:1":3342},{"statusCode":4,"data":5,"message":37},200,{"tenantId":6,"lang":7,"defaultLang":8,"siteUrl":9,"contactEmail":10,"brandName":11,"logoUrl":12,"siteName":11,"siteDescription":13,"ogImage":10,"robotsIndex":14,"socialLinks":10,"reservedSlugs":10,"seoPolicy":15},"stajic","sr","de","https:\u002F\u002Fstajic.de",null,"Stajic Platform","\u002FLogo_Planet.svg","Stajic Portal",true,{"branding":16,"relatedContent":17,"crossDomainLinks":18},{"logoUrl":12},{"enabled":14},[19,22,25,28,31,34],{"url":20,"label":21,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Ffigure.rocks","figure.rocks",{"url":23,"label":24,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Floving.rocks","loving.rocks",{"url":26,"label":27,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.com","bazify.com",{"url":29,"label":30,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.de","bazify.de",{"url":32,"label":33,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.at","bazify.at",{"url":35,"label":36,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.ba","bazify.ba","Portal settings resolved",[39,45],{"id":40,"name":41,"location":42,"isActive":14,"isDefault":43,"items":44},1,"main-navigation","header",false,[],{"id":46,"name":47,"location":48,"isActive":14,"isDefault":14,"items":49},4,"main-menu","sidebar",[50,66,79,93,103,118,133],{"id":51,"title":52,"url":60,"target":61,"icon":62,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":64,"portfolioId":10,"children":65},"item-18",{"de":53,"en":54,"es":55,"fr":56,"it":54,"ru":57,"sr":58,"zh":59},"Startseite","Home","Inicio","Accueil","Главная","Почетна","首页","\u002Ffull-stack-web-developer-munich-performance-seo-and-maintainable-builds","_self","i-lucide-home","page",111,[],{"id":67,"title":68,"url":75,"target":61,"icon":76,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":77,"portfolioId":10,"children":78},"item-22",{"de":69,"en":69,"es":70,"fr":69,"it":71,"ru":72,"sr":73,"zh":74},"Vision","Visión","Visione","Видение","Визија","想象","\u002Fueber-uns-webdesign-muenchen-webaplikation","i-lucide-eye",113,[],{"id":80,"title":81,"url":89,"target":61,"icon":90,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":91,"portfolioId":10,"children":92},"item-19",{"de":82,"en":83,"es":84,"fr":83,"it":85,"ru":86,"sr":87,"zh":88},"Leistungen","Services","Servicios","Servizi","Услуги","Услуге","服务","\u002Fservices-dienstleistungen-muenchen","i-lucide-wrench",116,[],{"id":94,"title":95,"url":99,"target":61,"icon":100,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":101,"portfolioId":10,"children":102},"item-23",{"de":96,"en":96,"es":96,"fr":96,"it":96,"ru":97,"sr":97,"zh":98},"Blog","Блог","博客","\u002Fblog","i-lucide-book-open",112,[],{"id":104,"title":105,"url":114,"target":61,"icon":115,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":116,"portfolioId":10,"children":117},"item-32",{"de":106,"en":107,"es":108,"fr":109,"it":110,"ru":111,"sr":112,"zh":113},"Neue Technologien","New Technologies","Nuevas tecnologías","Nouvelles technologies","Nuove tecnologie","Новые технологии","Нове технологије","新技术！","\u002Fneue-webtechnologien","i-lucide-sparkles",122,[],{"id":119,"title":120,"url":129,"target":61,"icon":130,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":131,"portfolioId":10,"children":132},"item-20",{"de":121,"en":122,"es":123,"fr":124,"it":125,"ru":126,"sr":127,"zh":128},"Kontakt","Contact us!","Contacto","Contact","Contatto","Контакт","Контактирајте нас","联系我们！","\u002Fcontact","i-lucide-mail",115,[],{"id":134,"title":135,"url":144,"target":61,"icon":145,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":146,"portfolioId":10,"children":147},"item-21",{"de":136,"en":137,"es":138,"fr":139,"it":140,"ru":141,"sr":142,"zh":143},"Unsere Arbeit","Our Work","Nuestro trabajo","Nos réalisations","I nostri lavori","Наши работы","Наши радови","文件夹","\u002Fportfolio","i-lucide-briefcase",114,[148,161,175,181,193],{"id":149,"title":150,"url":144,"target":61,"icon":159,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":146,"portfolioId":10,"children":160},"item-24",{"de":151,"en":152,"es":153,"fr":154,"it":155,"ru":156,"sr":157,"zh":158},"Alle Projekte","All Projects","Todos los proyectos","Tous les projets","Tutti i progetti","Все проекты","Сви пројекти","所有项目","i-lucide-grid-3x3",[],{"id":162,"title":163,"url":171,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":174},"item-29",{"de":164,"en":165,"es":166,"fr":167,"it":168,"ru":169,"sr":170,"zh":143},"Local Roots, Global Reach","Local Roots - Global Reach","Empresa local ","Entreprise locale","Azienda locale","Местная компания","Локално предузеће глобално тржиште","\u002Fportfolio\u002Flocal-roots-global-reach-communication-media-systems-for-modern-business","i-lucide-folder","custom",[],{"id":176,"title":177,"url":179,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":180},"item-28",{"de":178,"en":178,"es":178,"fr":178,"it":178,"ru":178,"sr":178,"zh":178},"Solr Suggester","\u002Fportfolio\u002Fsolr-fuzzy-suggester-und-solr-infix-suggester-abfrage-ueber-ajax-und-filterung",[],{"id":182,"title":183,"url":191,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":192},"item-27",{"de":184,"en":185,"es":186,"fr":187,"it":188,"ru":189,"sr":190,"zh":185},"Firmenwebseite SEO","Company Website SEO","Sitio web corporativo SEO","Site web d’entreprise SEO","Sito web aziendale SEO","Корпоративный сайт SEO","Пословна веб-страница SEO","\u002Fportfolio\u002Fseo-sem-branding-mobile-webseite-muenchen",[],{"id":194,"title":195,"url":203,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":204},"item-31",{"de":196,"en":197,"es":198,"fr":199,"it":200,"ru":201,"sr":202,"zh":197},"Digitalisierungsportal","Digitalization Portal","Portal de digitalización","Portail de numérisation","Portale di digitalizzazione","Портал цифровизации","Портал за дигитализацију","\u002Fportfolio\u002Fdigitalisierungsportal-archiv-museum-bibliothek-ead-lido-mets-mods",[],{"statusCode":4,"data":206,"message":3341},{"id":207,"title":208,"slug":209,"content":210,"contentJson":211,"excerpt":1539,"featuredImage":1540,"featuredImageAlt":1541,"featuredImageCaption":10,"featuredImageTitle":10,"featuredImageCopyright":10,"featuredImageAuthor":10,"featuredImageSourceUrl":10,"featuredImageLicense":10,"featuredImageIsAiGenerated":43,"status":1542,"publishedAt":1543,"createdAt":1544,"updatedAt":1545,"seoLocalePaths":1546,"categories":1555,"author":1568,"translations":1573},"489","Agentna AI objašnjena: Kada AI sistem može da planira, koristi alate i deluje","agentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act","\u003Cp>Agentna AI je AI sistem u kojem model može da teži cilju kroz više koraka tako što odlučuje šta će sledeće uraditi, koristeći alate ili druge sposobnosti, posmatrajući rezultate, ažurirajući svoje radno stanje i nastavljajući dok ne dostigne uslov za zaustavljanje. Sam model nije agent. Upotrebljiv agent takođe zahteva izvršno okruženje ili okvir koji upravlja kontekstom, izvršavanjem alata, stanjem, dozvolama, odobrenjima, greškama i petljom između odluka i posmatranja.\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--info my-6 rounded-xl border p-5 border-blue-300 bg-blue-50 dark:border-blue-900 dark:bg-blue-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">Direktan odgovor\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">Normalan poziv modela je obično \u003Cstrong>ulaz → model → izlaz\u003C\u002Fstrong>. Agentni sistem je bliži \u003Cstrong>cilj → odluka → alat\u002Fakcija → posmatranje → ažurirana odluka → … → rezultat\u003C\u002Fstrong>.\u003Cbr>\u003Cbr>Ključna razlika nije u tome da li aplikacija koristi LLM ili pozivanje funkcija. Već u tome da li sistem daje modelu značajnu kontrolu nad sledećim korakom višekoraknog procesa, dok izvršno okruženje ograničava šta model zapravo sme da radi.\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">Sposobnost nije ovlašćenje\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">Model može da zna kako da pozove alat. Izvršno okruženje može da izloži taj alat. Nijedna od tih činjenica ne znači da je trenutni korisnik ili agent ovlašćen da izvrši osnovnu poslovnu akciju. \u003Cstrong>Sposobnost alata, dozvola za alat i poslovno ovlašćenje su odvojeni slojevi.\u003C\u002Fstrong>\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">Napomena o aktuelnim izvorima — 8. oktobar 2026.\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">Terminologija agenata još uvek varira među dobavljačima i istraživačkim zajednicama. OpenAI trenutno definiše izvršna okruženja agenata oko višekoraknog rada, alata, stanja i orkestracije. Anthropic-ova praktična razlika ostaje korisna: radni tokovi prate unapred definisane putanje koda, dok agenti dinamički usmeravaju sopstveni proces i korišćenje alata. Ovaj članak stoga tretira „agentnu AI“ kao arhitektonski spektar, a ne kao jednu standardizovanu kategoriju proizvoda.\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Cnav class=\"editorjs-toc\" data-editorjs-toc=\"true\" aria-label=\"Sadržaj\">\u003Cstrong class=\"editorjs-toc__title\">Sadržaj\u003C\u002Fstrong>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-0\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-6\" class=\"editorjs-toc__link\">Šta agentna AI zaista znači\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-10\" class=\"editorjs-toc__link\">Najjednostavniji primer\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-15\" class=\"editorjs-toc__link\">Gde se jednostavan primer zaustavlja\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-18\" class=\"editorjs-toc__link\">Agent naspram radnog toka\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-21\" class=\"editorjs-toc__link\">Agentsko ponašanje je spektar, a ne binarna oznaka\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-24\" class=\"editorjs-toc__link\">Minimalna arhitektura agentskog sistema\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-26\" class=\"editorjs-toc__link\">Model nije agent\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-30\" class=\"editorjs-toc__link\">Korišćenje alata je centralno — ali samo korišćenje alata ne čini agenta\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-34\" class=\"editorjs-toc__link\">Sposobnost alata, dozvola i ovlašćenje su različiti\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-37\" class=\"editorjs-toc__link\">Izvršno okruženje ili okvir je stvarni sistem za izvršavanje\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-41\" class=\"editorjs-toc__link\">Planiranje je korisno, ali eksplicitni plan nije neophodan\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-44\" class=\"editorjs-toc__link\">Povratne informacije iz okruženja čine petlju korisnom\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-48\" class=\"editorjs-toc__link\">Stanje agenta nije isto kao kontekst modela\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-51\" class=\"editorjs-toc__link\">Memorija je opciona, nije definicija agenta\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-55\" class=\"editorjs-toc__link\">Inženjering konteksta postaje dinamičan u agentima\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-58\" class=\"editorjs-toc__link\">Alati za čitanje i alati sa sporednim efektima imaju različit rizik\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-60\" class=\"editorjs-toc__link\">Čovek u petlji je kontrolni mehanizam, a ne suprotnost agentnoj AI\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-64\" class=\"editorjs-toc__link\">Agenti zahtevaju eksplicitne uslove zaustavljanja\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-66\" class=\"editorjs-toc__link\">Oporavak je deo ponašanja agenta\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-70\" class=\"editorjs-toc__link\">Agentna AI ne zahteva više agenata\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-74\" class=\"editorjs-toc__link\">Agent protokoli su slojevi interoperabilnosti, a ne sam agent\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-78\" class=\"editorjs-toc__link\">Trajektorija je deo pouzdanosti agenta\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-83\" class=\"editorjs-toc__link\">Agentni sistemi povećavaju bezbednosnu površinu\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-85\" class=\"editorjs-toc__link\">Opservabilnost agenta mora pratiti petlju\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-88\" class=\"editorjs-toc__link\">Kako evaluirati agentni sistem\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-90\" class=\"editorjs-toc__link\">Kada je agent prikladan\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-93\" class=\"editorjs-toc__link\">Dokazi iz originalne implementacije\u003C\u002Fa>\u003Col class=\"editorjs-toc__list editorjs-toc__list--depth-1\">\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-94\" class=\"editorjs-toc__link\">Aaasaasa AI Client: model, izvršno okruženje i dozvole su odvojeni\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-99\" class=\"editorjs-toc__link\">Source of Truth Research Engine: ograničene agentske faze istraživanja\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-105\" class=\"editorjs-toc__link\">Uobičajeni načini neuspeha agentske AI\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-107\" class=\"editorjs-toc__link\">Uobičajene zablude\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-109\" class=\"editorjs-toc__link\">Praktičan redosled dizajna agenta\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-111\" class=\"editorjs-toc__link\">Kontrolna lista arhitekture agentne AI\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-113\" class=\"editorjs-toc__link\">Rubni slučajevi i ograničenja\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-119\" class=\"editorjs-toc__link\">Šta bi promenilo ovaj odgovor?\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-123\" class=\"editorjs-toc__link\">Povezano kanonsko znanje\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-127\" class=\"editorjs-toc__link\">Često postavljana pitanja\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-129\" class=\"editorjs-toc__link\">Pojmovnik\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-131\" class=\"editorjs-toc__link\">Zaključak\u003C\u002Fa>\u003C\u002Fli>\u003Cli class=\"editorjs-toc__item\">\u003Ca href=\"#section-135\" class=\"editorjs-toc__link\">Primarni izvori i aktuelne smernice\u003C\u002Fa>\u003C\u002Fli>\u003C\u002Fol>\u003C\u002Fnav>\n\u003Ch2 id=\"section-6\">Šta agentna AI zaista znači\u003C\u002Fh2>\n\u003Cp>Važan pomak od obične generativne AI ka agentnoj AI je kontrola nad procesom. Normalan asistent može da odgovori na pitanje koristeći kontekst koji dobije. Agent može da odluči da odgovaranje zahteva dodatne korake: pregledanje datoteke, pretragu repozitorijuma, upit ka API-ju, traženje pojašnjenja, pokretanje testa, ažuriranje tiketa, delegiranje podzadatka ili ponovni pokušaj nakon neuspele akcije.\u003C\u002Fp>\n\u003Cp>To ne zahteva neograničenu autonomiju. Agent može da radi unutar uskog sandbox-a, pod strogim dozvolama, uz odobrenje potrebno pre svake konsekventne akcije. Sistem je i dalje agentan ako model dinamički bira među dozvoljenim sledećim koracima.\u003C\u002Fp>\n\u003Cp>Arhitektura je stoga važnija od etikete. „Agent“ treba da opisuje ponašanje sistema: iterativno odlučivanje vođeno modelom nad alatima, stanjem i povratnim informacijama — a ne samo chatbot sa većim promptom.\u003C\u002Fp>\n\u003Ch2 id=\"section-10\">Najjednostavniji primer\u003C\u002Fh2>\n\u003Cp>Pretpostavimo da programer pita AI sistem: „Pronađi zašto test suite ne prolazi i ispravi grešku.“ Jedan poziv modela mogao bi samo da predloži verovatne uzroke na osnovu teksta koji je dobio.\u003C\u002Fp>\n\u003Cp>Agentni sistem za kodiranje može da pregleda repozitorijum, pronađe test koji ne prolazi, pročita relevantne datoteke, predloži izmenu, uredi kod, pokrene test, posmatra neuspeh, revidira implementaciju i ponovo pokrene test.\u003C\u002Fp>\n\u003Cp>Agentni deo nije samo to što postoje shell i datotečni alati. Već to što model može da koristi povratne informacije iz okruženja da izabere sledeći korak umesto da prati jedan potpuno unapred definisan niz.\u003C\u002Fp>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">Osnovna petlja agenta\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. Primanje cilja\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Korisnik ili nadređeni sistem definiše cilj i relevantna ograničenja.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. Izgradnja trenutnog konteksta\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Izvršno okruženje obezbeđuje instrukcije, stanje, istoriju, memoriju, alate i trenutne dokaze.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. Model odlučuje o sledećem koraku\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Model može da odgovori, pozove alat, zatraži informacije, delegira ili se zaustavi.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. Izvršno okruženje validira zahtev\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Dozvole, šeme, odobrenja i politika određuju da li predložena akcija sme da se izvrši.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. Izvršavanje alata ili akcije\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Spoljašnje okruženje se menja ili vraća nove informacije.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">6\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">6. Posmatranje rezultata\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Izvršno okruženje vraća strukturisani izlaz alata, greške ili promene stanja u sledeći korak modela.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">7\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">7. Nastavak ili zaustavljanje\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Petlja se ponavlja dok ne postigne uspeh, odbijanje, eskalaciju, ograničenje budžeta, istek vremena ili drugi uslov za zaustavljanje.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-15\">Gde se jednostavan primer zaustavlja\u003C\u002Fh2>\n\u003Cp>Nije svaki višekorakni AI sistem jednako agentan. Radni tok može da koristi nekoliko LLM poziva i alata, dok je svaki korak unapred određen u kodu. Drugi sistem može da dozvoli modelu da odluči koji alat će pozvati, kojim redosledom, koliko puta i kada će se zaustaviti.\u003C\u002Fp>\n\u003Cp>Oba mogu biti korisna. Razlika je u tome gde se nalazi kontrola. Unapred definisani radni tokovi stavljaju više kontrole u aplikacijski kod. Agenti premeštaju više taktičkih odluka o procesu u petlju model\u002Fizvršno okruženje.\u003C\u002Fp>\n\u003Ch2 id=\"section-18\">Agent naspram radnog toka\u003C\u002Fh2>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">Predefinisani tok rada i agentska kontrola\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">LLM tok rada\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">Agent\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">Put procesa\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">Sekvenca alata\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">Snaga\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">Rizik\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Cp>Anthropic eksplicitno razdvaja ova dva obrasca: tokovi rada orkestriraju modele i alate kroz predefinisane putanje koda, dok agenti dozvoljavaju modelima da dinamički usmeravaju sopstvene procese i korišćenje alata. Ovo nije jedina moguća terminologija, ali predstavlja korisnu arhitektonsku granicu.\u003C\u002Fp>\n\u003Ch2 id=\"section-21\">Agentsko ponašanje je spektar, a ne binarna oznaka\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Nivo\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Primer\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Ko odlučuje o sledećem koraku?\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Poziv jednog modela\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Sumiraj ovaj dokument\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Aplikacija poziva model jednom\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Odgovor uz pomoć alata\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model može koristiti web pretragu pre odgovaranja\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model bira iz ograničenih alata za jedan odgovor\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Strukturirani tok rada\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Klasifikuj → preuzmi → generiši → validiraj\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Tok rada aplikacije određuje faze\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Adaptivni tok rada\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model može birati između nekoliko grana i ponoviti\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Deljena kontrola između aplikacije i modela\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Agentska petlja\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model više puta bira alate\u002Fradnje na osnovu zapažanja\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model usmerava taktičko izvršavanje unutar ograničenja izvršnog okruženja\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Dugotrajni agent\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Agent pauzira, nastavlja, upravlja artefaktima i nastavlja dalje\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model + trajno izvršno okruženje upravljaju evoluirajućim izvršavanjem\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Cp>Nazivanje svakog sistema iznad „agentom“ može prikriti važne operativne razlike. Što je jača kontrola modela nad sekvencom, trajanjem i radnjama, to su važniji izolacija izvršnog okruženja, dozvole, praćenje, uslovi zaustavljanja i evaluacija trajektorije.\u003C\u002Fp>\n\u003Ch2 id=\"section-24\">Minimalna arhitektura agentskog sistema\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Komponenta\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Odgovornost\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Cilj \u002F zadatak\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Definiše šta sistem pokušava da postigne.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Tumači kontekst i odlučuje o sledećoj radnji ili izlazu.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Instrukcije\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Definišu ulogu, ograničenja, prioritete i politiku specifičnu za zadatak.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Sastavljač konteksta\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Gradi informacije vidljive modelu u svakom koraku.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Katalog alata\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Definiše mogućnosti koje model može zahtevati.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Izvršno okruženje \u002F okvir\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Pokreće petlju, izvršava alate, upravlja stanjem i rukuje uslovima zaustavljanja.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Sloj autorizacije\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Određuje da li je predložena radnja dozvoljena za trenutnog korisnika.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Stanje \u002F sesija\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Čuva napredak zadatka kroz različite faze ili korake izvršavanja.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Kanal zapažanja\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Vraća rezultate alata i promene okruženja u sledeći korak modela.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Odobrenja \u002F ljudska kontrola\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Pauzira radnje sa posledicama kada je potrebna provera.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Praćenje \u002F revizija\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Beleži pozive modela, alate, prelaze, odobrenja i neuspehe.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Evaluacija\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Meri ishode i trajektorije izvršavanja u odnosu na kriterijume prihvatanja.\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-26\">Model nije agent\u003C\u002Fh2>\n\u003Cp>Jezički model proizvodi izlaze na osnovu ulaza. On sam po sebi ne poseduje fajl sistem, ne izvršava shell komandu, ne održava trajno stanje zadatka, ne sprovodi dozvole niti se automatski ponovo poziva.\u003C\u002Fp>\n\u003Cp>Te sposobnosti dolaze iz okolnog izvršnog okruženja. Isti model može se ponašati kao jednostavan chat model u jednoj aplikaciji i kao motor za odlučivanje unutar agentske petlje u drugoj.\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--success my-6 rounded-xl border p-5 border-emerald-300 bg-emerald-50 dark:border-emerald-900 dark:bg-emerald-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">Arhitektonsko pravilo\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">\u003Cstrong>Sposobnost modela određuje koje odluke mogu biti predložene. Arhitektura izvršnog okruženja određuje šta se zapravo može dogoditi.\u003C\u002Fstrong>\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-30\">Korišćenje alata je centralno — ali samo korišćenje alata ne čini agenta\u003C\u002Fh2>\n\u003Cp>Alati omogućavaju modelu da pribavi informacije i utiče na spoljne sisteme. Primeri uključuju čitanje baze podataka, operacije sa fajlovima, izvršavanje shell komandi, web pretragu, kontrolu pregledača, API pozive, ažuriranje tiketa ili delegiranje specijalizovanim agentima.\u003C\u002Fp>\n\u003Cp>Poziv jednog modela može koristiti jedan alat i dalje ostati ograničen odgovor uz pomoć alata, a ne dugotrajni agent. Agentsko ponašanje se pojavljuje kada zapažanja alata hrane adaptivnu petlju u kojoj model bira šta će sledeće uraditi.\u003C\u002Fp>\n\u003Cp>Dizajn alata je važan jer su alati ugovor između rezonovanja modela i spoljne stvarnosti. Dvosmisleni ili preklapajući alati stvaraju greške u rutiranju; veliki nestrukturirani izlazi zagađuju kontekst; alati sa širokim sporednim efektima povećavaju obim štete.\u003C\u002Fp>\n\u003Ch2 id=\"section-34\">Sposobnost alata, dozvola i ovlašćenje su različiti\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Sloj\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Pitanje\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Sposobnost\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Može li ovo izvršno okruženje tehnički izvesti operaciju?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Izloženost alata\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li je ta sposobnost dostupna ovom agentu?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Dozvola\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Sme li ovaj agent\u002Fsesija da je koristi pod trenutnom politikom?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Korisnička autorizacija\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li je korisnik koji zahteva dozvoljen da izazove ovu operaciju?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Poslovno ovlašćenje\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li je operacija validna prema pravilima domena, odobrenjima i ograničenjima?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Izvršavanje\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li se operacija zaista dogodila?\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Revizija\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Može li sistem dokazati ko je zahtevao, odobrio i izvršio operaciju?\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Cp>Ovi slojevi se često spajaju u prototipovima. Model vidi alat za refundaciju i stoga izgleda da može izvršiti refundacije. U produkciji, alat bi ipak trebalo da validira nalog, korisnika, transakciju, iznos, politiku i uslove odobrenja nezavisno od zahteva modela.\u003C\u002Fp>\n\u003Ch2 id=\"section-37\">Izvršno okruženje ili okvir je stvarni sistem za izvršavanje\u003C\u002Fh2>\n\u003Cp>OpenAI-jeva trenutna dokumentacija o agentima eksplicitno pravi razliku između izvršnih okruženja. Različita izvršna okruženja mogu upravljati orkestracijom, stanjem, alatima, sandbox-ovima i izvršavanjem na različitim mestima, dok model ostaje samo jedan deo sistema.\u003C\u002Fp>\n\u003Cp>Agents SDK opisuje petlju koja više puta poziva trenutni model, pregleda izlaz, izvršava tražene alate ili predaje, i nastavlja dok model ne vrati konačan odgovor ili drugu stvarnu tačku zaustavljanja.\u003C\u002Fp>\n\u003Cp>To znači da odluke o arhitekturi agenta uključuju gde se izvršava orkestracija, gde se čuva stanje, ko izvršava alate, koji sandbox sadrži neželjene efekte i ko upravlja ponovnim pokušajima, vremenskim ograničenjima i mogućnošću nastavljanja.\u003C\u002Fp>\n\u003Ch2 id=\"section-41\">Planiranje je korisno, ali eksplicitni plan nije neophodan\u003C\u002Fh2>\n\u003Cp>Agenti se često opisuju kao sistemi koji „planiraju“. U praksi, planiranje može biti eksplicitno ili implicitno. Agent može prvo da proizvede vidljiv višekoračni plan, ili može da bira jednu sledeću akciju po jednu i revidira nakon svakog zapažanja.\u003C\u002Fp>\n\u003Cp>Za veoma neizvesne zadatke, planiranje kratkog horizonta može biti sigurnije jer okruženje može poništiti dugoročni plan. Arhitektonski zahtev je sposobnost izbora i revizije akcija na osnovu cilja, trenutnog stanja i novih dokaza.\u003C\u002Fp>\n\u003Ch2 id=\"section-44\">Povratne informacije iz okruženja čine petlju korisnom\u003C\u002Fh2>\n\u003Cp>Agent postaje operativno značajan kada može da primeti da li je njegova akcija uspela. Izlaz alata, rezultati testova, odgovori API-ja, stanje fajl sistema, stanje pregledača i zapisi aplikacija pružaju spoljne dokaze koje sistem može da koristi za reviziju sledeće odluke.\u003C\u002Fp>\n\u003Cp>Anthropic-ove smernice za agente naglašavaju ovu povratnu petlju: agenti koriste alate, dobijaju osnovnu istinu iz okruženja, procenjuju napredak i nastavljaju ili traže ljudski unos.\u003C\u002Fp>\n\u003Caside class=\"editorjs-callout editorjs-callout--warning my-6 rounded-xl border p-5 border-amber-300 bg-amber-50 dark:border-amber-900 dark:bg-amber-950\u002F20\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">Samoizveštavanje nije dokaz iz okruženja\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">To što agent kaže „zadatak je završen“ ne dokazuje završetak. Gde je moguće, proverite konačno stanje kroz spoljni sistem, test, fajl, zapis transakcije ili drugi vidljivi ishod.\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-48\">Stanje agenta nije isto kao kontekst modela\u003C\u002Fh2>\n\u003Cp>Dugotrajan zadatak može zahtevati stanje koje ne može ili ne bi trebalo da ostane u kontekstu modela: ID-ovi zadataka, kontrolne tačke, artefakti, odobrenja, identifikatori spoljnih objekata, brojači ponovnih pokušaja i status radnog toka.\u003C\u002Fp>\n\u003Cp>Izvršno okruženje može da sačuva ovo trajno stanje izvan prozora modela i rekonstruiše kontekst potreban za sledeći korak. To održava kontekst vidljiv modelu fokusiranim dok se održava kontinuitet i mogućnost nastavljanja.\u003C\u002Fp>\n\u003Ch2 id=\"section-51\">Memorija je opciona, nije definicija agenta\u003C\u002Fh2>\n\u003Cp>Agent može uspešno da radi bez dugoročne memorije ako kompletan zadatak staje unutar jednog ograničenog izvršavanja. Memorija postaje korisna kada informacije moraju da se zadrže kroz sesije, zadatke ili duge horizonte izvršavanja.\u003C\u002Fp>\n\u003Cp>RAG, memorija, stanje i kontekst rešavaju različite probleme. Tretiranje vektorske baze podataka kao „memorije agenta“ ili istorije razgovora kao „mašine stanja“ obično skriva važne granice životnog ciklusa i autoriteta.\u003C\u002Fp>\n\u003Caside class=\"editorjs-referral my-6\">\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fsr\u002Fblog\u002Fai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context\" class=\"flex flex-col sm:flex-row gap-4 rounded-xl border border-gray-200 dark:border-gray-700 p-4 transition hover:border-primary-500\">\u003Cdiv class=\"min-w-0 flex-1\">\u003Cstrong class=\"block text-lg text-gray-900 dark:text-gray-100\">Memorija AI agenta nije RAG: Kako razdvojiti memoriju, pretragu, stanje i kontekst\u003C\u002Fstrong>\u003Cp class=\"mt-2 text-sm text-gray-600 dark:text-gray-300\">Praktična arhitektura koja razdvaja trajnu memoriju, autoritativno stanje aplikacije, pretragu i kontekst koji se dostavlja modelu.\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">Pročitajte članak o arhitekturi memorije →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Ch2 id=\"section-55\">Inženjering konteksta postaje dinamičan u agentima\u003C\u002Fh2>\n\u003Cp>Svaki poziv alata može da proizvede novi kontekst. Svaki korak takođe može da učini ranije informacije zastarelim. Snažno izvršno okruženje agenta zato ponovo gradi ili oblikuje kontekst kako izvršavanje napreduje, umesto da iznova prikazuje sve neograničeno.\u003C\u002Fp>\n\u003Cp>Definicije alata, stanje zadatka, preuzeti dokazi, zapažanja i memorija svi se takmiče za pažnju modela. Dugotrajni agenti zahtevaju skraćivanje, sažimanje ili učitavanje na vreme kako bi kontekst ostao relevantan za trenutnu odluku.\u003C\u002Fp>\n\u003Ch2 id=\"section-58\">Alati za čitanje i alati sa sporednim efektima imaju različit rizik\u003C\u002Fh2>\n\u003Csection class=\"editorjs-comparison my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">Pristup informacijama naspram spoljne akcije\u003C\u002Fh3>\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left dark:border-gray-700 dark:bg-gray-900\">\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">Čitanje \u002F posmatranje\u003C\u002Fth>\u003Cth class=\"border border-gray-300 bg-gray-50 px-4 py-3 text-left font-semibold dark:border-gray-700 dark:bg-gray-900\">Pisanje \u002F delovanje\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">Primeri\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">Glavni rizik\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-3 text-left font-semibold dark:border-gray-700\">Tipična kontrola\u003C\u002Fth>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-3 dark:border-gray-700\">\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-60\">Čovek u petlji je kontrolni mehanizam, a ne suprotnost agentnoj AI\u003C\u002Fh2>\n\u003Cp>Agent ne prestaje da bude agentan zato što čovek odobrava konsekventne korake. Model i dalje može autonomno da pregleda, zaključuje, pretražuje i priprema akciju dok izvršno okruženje zahteva ljudsku potvrdu pre izvršenja.\u003C\u002Fp>\n\u003Cp>OpenAI-ove trenutne smernice za bezbednost agenata eksplicitno preporučuju odobrenja za operacije alata u rizičnijim radnim tokovima. Anthropic takođe naglašava kontrolne tačke i ljudsku procenu tamo gde agenti naiđu na prepreke ili konsekventne odluke.\u003C\u002Fp>\n\u003Cp>Korisno arhitektonsko pitanje nije „čovek ili autonomno?“ već koje odluke mogu da se delegiraju, koje zahtevaju pregled i koje moraju da ostanu determinističke?\u003C\u002Fp>\n\u003Ch2 id=\"section-64\">Agenti zahtevaju eksplicitne uslove zaustavljanja\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Uslov zaustavljanja\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Svrha\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Uspešan verifikovan ishod\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Završi kada je spoljno ciljno stanje potvrđeno.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Maksimalan broj koraka\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Spreči nekontrolisane petlje.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Vremenski budžet\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ograniči izvršavanje po stvarnom vremenu.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Budžet troškova\u002Ftokena\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ograniči potrošnju resursa.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Detektor ponovljene akcije\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Zaustavi petlje koje više ne napreduju.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Granica dozvola\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Pauziraj ili zaustavi kada sledeća potrebna akcija nije dozvoljena.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Kontrolna tačka ljudskog odobrenja\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Sačekaj pre konsekventnog izvršenja.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Neopoziv kvar alata\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Eskaliraj umesto beskonačnog ponavljanja.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Prag neizvesnosti\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Zatraži pojašnjenje kada zadatak ne može bezbedno da se zaključi.\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-66\">Oporavak je deo ponašanja agenta\u003C\u002Fh2>\n\u003Cp>Agenti rade u okruženjima koja otkazuju: API-ji isteknu, datoteke se promene, akreditivi isteknu, web stranice se premeste i alati vraćaju neispravan izlaz. Koristan agentni sistem zato zahteva ponašanje oporavka, a ne samo petlju alata za idealan scenario.\u003C\u002Fp>\n\u003Cp>Oporavak može da uključi ponovni pokušaj sa ograničenjima, izbor drugog alata, ponovno čitanje trenutnog stanja, pitanje korisniku, vraćanje delimične akcije ili eskalaciju ka čoveku.\u003C\u002Fp>\n\u003Cp>Ponovni pokušaji takođe zahtevaju svest o idempotentnosti. Ponavljanje čitanja je obično niskog rizika; ponavljanje plaćanja ili slanja poruke može da stvori duplikate sporednih efekata.\u003C\u002Fp>\n\u003Ch2 id=\"section-70\">Agentna AI ne zahteva više agenata\u003C\u002Fh2>\n\u003Cp>Jedan agent sa jasnim skupom alata je često jednostavniji i lakši za evaluaciju od multi-agent arhitekture. Više agenata je korisno kada specijalizacija materijalno poboljšava izolaciju alata, izolaciju politika, jasnoću upita, vlasništvo ili čitljivost traga.\u003C\u002Fp>\n\u003Cp>OpenAI-ove trenutne smernice za orkestraciju eksplicitno preporučuju početak sa jednim agentom gde je to moguće i dodavanje specijalista samo kada se ugovor ili granica vlasništva materijalno promene.\u003C\u002Fp>\n\u003Cp>Multi-agentni sistemi dodaju nove probleme: kvalitet delegiranja, duplirani kontekst, konfliktno stanje, semantiku predaje, identitet, troškove i distribuirano rukovanje greškama.\u003C\u002Fp>\n\u003Ch2 id=\"section-74\">Agent protokoli su slojevi interoperabilnosti, a ne sam agent\u003C\u002Fh2>\n\u003Cp>Protokoli kao što su MCP i A2A mogu učiniti arhitekturu agenta interoperabilnom, ali sami po sebi ne stvaraju petlju agenta. MCP može izložiti alate i resurse. A2A može povezati nezavisno implementirane agente. Aplikaciji je i dalje potrebno izvršno okruženje, autorizacija, stanje, evaluacija i domenska logika.\u003C\u002Fp>\n\u003Cp>Zato sposobnost protokola mora ostati odvojena od poslovnog ovlašćenja. Otkrivanje alata putem MCP-a ne dokazuje da trenutni principal sme da ga koristi. Prijem zadatka putem A2A ne dokazuje da udaljeni agent sme da izvrši svaku traženu radnju.\u003C\u002Fp>\n\u003Caside class=\"editorjs-referral my-6\">\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fde\u002Fblog\u002Fmcp-vs-a2a-vs-ucp-vs-ap2-vs-a2ui-the-agent-protocol-stack-explained\" class=\"flex flex-col sm:flex-row gap-4 rounded-xl border border-gray-200 dark:border-gray-700 p-4 transition hover:border-primary-500\">\u003Cdiv class=\"min-w-0 flex-1\">\u003Cstrong class=\"block text-lg text-gray-900 dark:text-gray-100\">MCP vs A2A vs UCP vs AP2 vs A2UI: Objašnjen stek agent protokola\u003C\u002Fstrong>\u003Cp class=\"mt-2 text-sm text-gray-600 dark:text-gray-300\">Mapa odgovornosti protokola koja pokazuje zašto pristup alatima, saradnja agenata, trgovina, ovlašćenje plaćanja i UI vođen agentom pripadaju različitim granicama interoperabilnosti.\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">Pročitajte stek agent protokola →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Ch2 id=\"section-78\">Trajektorija je deo pouzdanosti agenta\u003C\u002Fh2>\n\u003Cp>Konačan odgovor je nedovoljan dokaz za agentni sistem jer agent može doći do ispravnog rezultata preko nesigurne ili nevalidne putanje. Može koristiti neovlašćeni alat, preskočiti obaveznu proveru, ponoviti sporedni efekat, osloniti se na zastarelo stanje ili slučajno uspeti.\u003C\u002Fp>\n\u003Cp>Evaluacija zato zahteva tragove izvršavanja: odluke, pozive alata, odobrenja, opservacije, promene stanja i konačan ishod. Trenutne OpenAI smernice za bezbednost preporučuju ocenjivače tragova i evaluacije; Anthropic-ove smernice za evaluaciju agenata iz 2026. takođe tretiraju višekratne trajektorije alata kao evaluacione objekte prvog reda.\u003C\u002Fp>\n\u003Cp>Jače pitanje pouzdanosti je: Da li je agent postigao prihvatljiv ishod kroz prihvatljivu, obnovljivu i revizibilnu trajektoriju?\u003C\u002Fp>\n\u003Caside class=\"editorjs-referral my-6\">\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fsr\u002Fblog\u002Fai-agent-reliability-why-the-final-answer-is-not-enough\" class=\"flex flex-col sm:flex-row gap-4 rounded-xl border border-gray-200 dark:border-gray-700 p-4 transition hover:border-primary-500\">\u003Cdiv class=\"min-w-0 flex-1\">\u003Cstrong class=\"block text-lg text-gray-900 dark:text-gray-100\">Pouzdanost AI agenata: Zašto konačan odgovor nije dovoljan\u003C\u002Fstrong>\u003Cp class=\"mt-2 text-sm text-gray-600 dark:text-gray-300\">Zašto proizvodna evaluacija mora da ispituje trajektorije, korišćenje alata, prelaze stanja i obnovljivost, a ne samo konačne odgovore.\u003C\u002Fp>\u003Cspan class=\"mt-3 inline-flex text-sm font-medium text-primary-600 dark:text-primary-400\">Pročitajte članak o pouzdanosti →\u003C\u002Fspan>\u003C\u002Fdiv>\u003C\u002Fa>\u003C\u002Faside>\n\u003Ch2 id=\"section-83\">Agentni sistemi povećavaju bezbednosnu površinu\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Rizik\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Zašto ga agenti pojačavaju\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Arhitektonski odgovor\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ubacivanje upita\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Nepouzdan sadržaj može uticati na buduće odluke o alatima\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Odvojite instrukcije od podataka; ograničite alate; sanitizujte ili strukturirajte spoljni unos gde je moguće\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Prekomerna ovlašćenja\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Greške u rezonovanju mogu postati stvarni sporedni efekti\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Najmanja privilegija, ograničeni akreditivi, politika po alatu i odobrenja\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Izlaganje akreditiva\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Alatima mogu biti potrebne moćne tajne\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Držite tajne van konteksta modela; posredujte pristup kroz pouzdano izvršno okruženje\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Zbunjeni posrednik\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Agent može delovati sa ovlašćenjem širim od korisnika koji zahteva\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Vežite izvršavanje za identitet korisnika\u002Fservisa i ponovo autorizujte posledične radnje\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Petlje bez kontrole\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model više puta poziva alate bez napretka\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Budžeti za korake, vreme i troškove plus detekcija petlji\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Pomeranje stanja\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Okruženje se menja nakon što agent formira plan\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ponovo pročitajte autoritativno stanje pre posledičnih radnji\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Indirektno ubacivanje\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Sadržaj alata\u002Fveba\u002Fdokumenta sadrži instrukcije usmerene na model\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Tretirajte spoljni sadržaj kao nepouzdane podatke, a ne kao instrukcijsko ovlašćenje\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Praznina u reviziji\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Konačan rezultat ne može pokazati šta je izvršeno\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Pratite pozive alata, odobrenja, identitete i promene stanja\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-85\">Opservabilnost agenta mora pratiti petlju\u003C\u002Fh2>\n\u003Cp>Tradicionalna opservabilnost servisa beleži zahteve, latenciju i greške. Opservabilnost agenta zahteva dodatni model izvršavanja: koji agent je bio aktivan, koja verzija modela je donela odluku, koji kontekst je bio dostupan, koji alat je izabran, koji argumenti su poslati, koji rezultat je vraćen i zašto je izvršavanje zaustavljeno.\u003C\u002Fp>\n\u003Cp>Za osetljive sisteme, sami tragovi zahtevaju kontrolu pristupa i politiku zadržavanja jer upiti, izlazi alata i artefakti mogu sadržati poverljive podatke.\u003C\u002Fp>\n\u003Ch2 id=\"section-88\">Kako evaluirati agentni sistem\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Dimenzija\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Pitanje\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Primer dokaza\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Uspeh zadatka\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li se traženi ishod dogodio?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Spoljno stanje, testovi, poslovni ishod\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Kvalitet trajektorije\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li su koraci bili prihvatljivi?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Trag alata\u002Fradnji\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Izbor alata\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li je agent izabrao odgovarajuće sposobnosti?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Očekivani vs stvarni pozivi alata\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Poštovanje dozvola\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li je ostao unutar dozvoljenog ovlašćenja?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Dnevnici autorizacije i testovi odbijenih radnji\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Rukovanje stanjem\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li je koristio trenutno autoritativno stanje?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Provere svežine i testovi promene stanja\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Oporavak\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li je ispravno odgovorio na greške?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ubačeni scenariji tajmauta\u002Fgreške\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ponašanje zaustavljanja\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li se zaustavio na pravom mestu?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Brojevi koraka, detekcija petlji, dokaz konačnog stanja\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ljudska eskalacija\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li je pitao kada je pregled bio potreban?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Tragovi odobrenja\u002Feskalacije\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Trošak\u002Flatencija\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li je autonomija vredela operativnog troška?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Tokeni, pozivi alata, trajanje\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Robusnost\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Da li preživljava realistične varijacije okruženja?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ponovljeni i adversarni pokušaji\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-90\">Kada je agent prikladan\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Koristite agenta kada\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Preferirajte radni tok ili jednostavan poziv kada\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Broj ili redosled koraka ne može biti pouzdano poznat unapred\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Sekvenca je stabilna i deterministička\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Sistem mora da pregleda okruženje i prilagodi se\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Jedan korak preuzimanja i generisanja je dovoljan\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Nekoliko alata može biti korisno u zavisnosti od međurezultata\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Jedan poznati API poziv rešava zadatak\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Zadatak ima koristi od iterativne verifikacije ili ispravke\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Odgovor može biti proizveden direktno iz dostavljenog konteksta\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Neuspesi zahtevaju fleksibilno ponašanje oporavka\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Grane neuspeha su jednostavne i mogu se eksplicitno kodirati\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ljudski pregled može biti umetnut na smislenim kontrolnim tačkama\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Svaki korak je visokorizičan i mora se ionako ručno kontrolisati\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Očekivana vrednost opravdava dodatnu latenciju, trošak i složenost\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Predvidivost i niska cena su važniji od fleksibilnosti\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Cp>Snažan podrazumevani pristup je započeti sa najjednostavnijim rešenjem koje funkcioniše i povećavati agentsku složenost samo kada fleksibilnost proizvodi merljivu vrednost. Agenti razmenjuju predvidivost, latenciju i trošak za adaptivno izvršavanje.\u003C\u002Fp>\n\u003Ch2 id=\"section-93\">Dokazi iz originalne implementacije\u003C\u002Fh2>\n\u003Ch3 id=\"section-94\">Aaasaasa AI Client: model, izvršno okruženje i dozvole su odvojeni\u003C\u002Fh3>\n\u003Cp>Aaasaasa AI Client eksplicitno razdvaja agenta\u002Fklijenta, provajdera, model, lokaciju izvršnog okruženja i dozvole. Njegova arhitektonska dokumentacija tretira dozvole kao centralnu politiku alata\u002Fradnog prostora, a ne kao osobinu modela.\u003C\u002Fp>\n\u003Cp>Ista aplikacija može da izloži Direct Chat bez alata za fajl sistem ili ljusku, dok Codex izvršno okruženje radi pod izabranim radnim prostorom i profilom dozvola. Ovo demonstrira ključnu granicu agentske arhitekture: promena površine izvršnog okruženja\u002Falata menja šta sistem može da uradi čak i kada pristup modelu ostaje dostupan.\u003C\u002Fp>\n\u003Cp>Repozitorijum takođe razlikuje lokalno Codex izvršno okruženje od lokacije modela: lokalno izvršno okruženje može pozvati model u oblaku. Ovo sprečava uobičajenu grešku izjednačavanja „agent se izvršava lokalno“ sa „inferencija je lokalna“.\u003C\u002Fp>\n\u003Cp>Implementacija onemogućava ugrađene putanje izvršavanja čija semantika odobravanja ne zadovoljava zahtevani model dozvola. Ovo podržava princip da sposobnost agenta ne treba da zaobiđe autorizaciju izvršnog okruženja samo zato što osnovni okvir može da izvršava alate.\u003C\u002Fp>\n\u003Ch3 id=\"section-99\">Source of Truth Research Engine: ograničene agentske faze istraživanja\u003C\u002Fh3>\n\u003Cp>Source of Truth Research Engine koristi ograničen istraživački tok: otkrij → pribavi → izvuci → verifikuj → protivreči → sintetiši. Istraživački poslovi mogu se izvršavati kroz AI izvršno okruženje dok dokazi, izvori, tvrdnje i protivrečnosti ostaju u eksternom trajnom skladištu.\u003C\u002Fp>\n\u003Cp>Ovo je namerno kontrolisanije od neograničenog autonomnog istraživačkog agenta. Faze pružaju zaštitne ograde oko toga koja vrsta posla treba sledeće da se obavi, dok i dalje dozvoljavaju istraživanje vođeno modelom unutar svakog ograničenog zadatka.\u003C\u002Fp>\n\u003Cp>Ta razlika je koristan dokaz za dizajn agenata: autonomija može biti smeštena unutar strukturisanog okvira isporuke, umesto da se primenjuje uniformno na ceo proces.\u003C\u002Fp>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Implementirani obrazac\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Lekcija agentske arhitekture\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Direct Chat nema OS alate\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model može postojati bez agentske sposobnosti izvršavanja.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Codex izvršno okruženje ima profil dozvola radnog prostora\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ovlašćenje alata pripada politici izvršnog okruženja, a ne sposobnosti modela.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Provajder\u002Fmodel\u002Fizvršno okruženje su odvojeni koncepti\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Lokacija agentskog okvira i lokacija inferencije su nezavisne odluke.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Broker dozvola za izvršna okruženja sa alatima\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Izlaganje sposobnosti može biti centralizovano i upravljano.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ograničene faze istraživanja\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Autonomija može da radi unutar eksplicitnih granica procesa.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Trajne tvrdnje\u002Fdokazi izvan konteksta modela\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Stanje agenta i dokazi ne moraju da žive samo u istoriji razgovora.\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Caside class=\"editorjs-callout editorjs-callout--note my-6 rounded-xl border p-5 border-gray-300 bg-gray-50 dark:border-gray-700 dark:bg-gray-900\u002F40\" role=\"note\">\u003Cstrong class=\"block mb-2 text-gray-900 dark:text-gray-100\">Granica dokaza\u003C\u002Fstrong>\u003Cdiv class=\"text-gray-700 dark:text-gray-200\">Ovi projekti demonstriraju konkretne obrasce agenta\u002Fizvršnog okruženja, dozvola i ograničenog istraživanja. Nisu predstavljeni kao dokaz masovnog komercijalnog uvođenja autonomnih agenata.\u003C\u002Fdiv>\u003C\u002Faside>\n\u003Ch2 id=\"section-105\">Uobičajeni načini neuspeha agentske AI\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Način neuspeha\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Šta je zapravo zakazalo\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„Agent“ je samo chatbot sa alatima navedenim u upitu\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ne postoji pouzdana petlja izvršnog okruženja ili arhitektura izvršavanja alata\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Podrška za alate se tretira kao dozvola\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Granice sposobnosti i autorizacije su srušene\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Agent veruje sopstvenoj izjavi o završetku\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ishod nije verifikovan u odnosu na eksterno stanje\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Svaki zadatak postaje multi-agent\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Složenost raste bez stvarne granice vlasništva ili specijalizacije\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Istorija razgovora se koristi kao trajno stanje\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Mogućnost nastavka i autoritativno stanje postaju krhki\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Agent slepo ponavlja sporedne efekte\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Duplikati poruka, plaćanja ili promena stanja postaju mogući\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Nema ograničenja koraka\u002Ftroškova\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Agent može da se vrti u nedogled ili troši nekontrolisane resurse\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Izlaz alata se veruje kao instrukcija\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Indirektna injekcija upita može preusmeriti ponašanje\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Tačan konačni odgovor je jedina evaluacija\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Nesigurne ili nevažeće putanje ostaju nevidljive\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Nadogradnja modela se tretira kao transparentna\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Izbor alata, planiranje i ponašanje zaustavljanja mogu se promeniti\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Jedan širok alat izlaže mnoge privilegovane operacije\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Radijus eksplozije se povećava i nameru je teže validirati\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ljudsko odobrenje postoji, ali recenzent nema kontekst\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Odobrenje postaje ceremonijalno, a ne efektivno\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-107\">Uobičajene zablude\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Zabluda\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Ispravka\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„LLM je agent.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Model je komponenta odlučivanja; agent je okolni sistem koji upravlja alatima, stanjem i iteracijom.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„Pozivanje alata automatski znači agentska AI.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Jedan ograničen poziv alata možda ne uključuje adaptivnu agentsku petlju sa više koraka.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„Agenti moraju biti potpuno autonomni.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Agentski sistemi mogu zahtevati odobrenja i raditi unutar uskih granica dozvola.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„Agenti moraju imati dugoročnu memoriju.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Memorija je opciona; mnogi korisni agenti završavaju ograničene zadatke bez memorije između sesija.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„Agenti moraju prvo napraviti pisani plan.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Planiranje može biti eksplicitno ili implicitno i može se odvijati korak po korak.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„Multi-agent je napredniji od single-agent.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Složeniji je; koristite ga samo kada specijalizacija ili granice vlasništva to opravdavaju.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„MCP stvara agenta.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">MCP izlaže alate\u002Fresurse; izvršno okruženje i dalje zahteva agentsku petlju i model autorizacije.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„Lokalno izvršno okruženje znači da je model lokalni.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Lokacija izvršnog okruženja i lokacija inferencije\u002Fprovajdera su odvojene.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„Ako je konačni rezultat tačan, agent je radio ispravno.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Nesigurna ili neovlašćena putanja i dalje može proizvesti tačan rezultat.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">„Ljudsko odobrenje uklanja autonomiju.“\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Odobrenje može ograničiti izabrane akcije dok ostatak procesa ostaje vođen modelom.\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-109\">Praktičan redosled dizajna agenta\u003C\u002Fh2>\n\u003Csection class=\"editorjs-process my-6\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">Dizajnirajte agenta od ovlašćenja ka spolja\u003C\u002Fh3>\u003Cdiv class=\"grid grid-cols-1 md:grid-cols-2 xl:grid-cols-3 gap-4\">\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">1\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">1. Definišite ishod\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Navedite koji spoljni rezultat ili artefakt dokazuje uspeh zadatka.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">2\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">2. Odlučite da li je agent zaista potreban\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Preferirajte jednostavan poziv ili deterministički tok rada kada je put predvidiv.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">3\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">3. Identifikujte stanje i izvor istine\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Definišite koji sistemi poseduju trenutne činjenice, napredak zadatka i poslovno stanje.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">4\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">4. Definišite površinu alata\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Izložite najmanji skup jasnih sposobnosti potrebnih za zadatak.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">5\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">5. Povežite identitet i dozvole\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Razdvojite korisničko ovlašćenje, dozvole agenta\u002Fruntime-a i sposobnosti alata.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">6\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">6. Izaberite granice autonomije\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Navedite šta model može dinamički da odlučuje, a šta ostaje deterministički.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">7\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">7. Dodajte kontrolne tačke za odobrenje\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Zahtevajte pregled pre posledičnih ili nepovratnih radnji gde je to prikladno.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">8\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">8. Definišite zaustavljanje i oporavak\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Postavite dokaz uspeha, budžete, tajmaute, ponovne pokušaje, eskalaciju i kontrolu petlji.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">9\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">9. Dizajnirajte upravljanje kontekstom\u002Fstanjem\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Držite trenutno stanje, memoriju, zapažanja alata i trajne artefakte u ispravnim slojevima.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">10\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">10. Pratite putanju\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Zabeležite dovoljno strukture izvršavanja za otklanjanje grešaka i reviziju odluka modela\u002Falata.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">11\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">11. Procenite realistične neuspehe\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Testirajte zastarelo stanje, greške alata, ubacivanje upita, dvosmislene zahteve i promenjena okruženja.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv class=\"editorjs-process__step min-w-0  rounded-xl border border-gray-200 dark:border-gray-700 p-4\">\u003Cdiv class=\"text-xs font-semibold text-gray-500 dark:text-gray-400\">12\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 font-semibold text-gray-900 dark:text-gray-100\">12. Proširite autonomiju samo na osnovu dokaza\u003C\u002Fdiv>\u003Cdiv class=\"mt-1 text-sm text-gray-600 dark:text-gray-300\">Povećajte dozvole ili horizont izvršavanja kada evaluacija pokaže da korist opravdava rizik.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-111\">Kontrolna lista arhitekture agentne AI\u003C\u002Fh2>\n\u003Cdiv class=\"overflow-x-auto\">\u003Ctable class=\"w-full border-collapse\">\u003Cthead>\u003Ctr>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Pitanje\u003C\u002Fth>\u003Cth class=\"border border-gray-300 px-4 py-2 text-left font-semibold\">Očekivani dokaz\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Šta dokazuje uspeh?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Spoljni ishod, artefakt, test ili autoritativno stanje.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Zašto je agent potreban?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Put zaista zavisi od posrednih zapažanja.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Koje odluke pokreće model?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Eksplicitna granica autonomije.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Koji alati postoje?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Mali, dokumentovan, nedvosmislen skup sposobnosti.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ko može da koristi svaki alat?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Politika autorizacije svesna identiteta i konteksta.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Koje radnje zahtevaju odobrenje?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Pravila pregleda zasnovana na posledicama.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Gde se čuva stanje zadatka?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Stanje u vlasništvu aplikacije, odvojeno od prolaznog konteksta modela.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Kako se agent oporavlja?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Ponašanje ponovnog pokušaja, ponovnog čitanja, vraćanja, pojašnjenja i eskalacije.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Kako se zaustavlja?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Verifikovani završetak plus ograničenja koraka\u002Fvremena\u002Ftroškova.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Kako su zaštićeni sporedni efekti?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Validacija, idempotentnost, najmanja privilegija i potvrda.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Može li se izvršavanje rekonstruisati?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Tragovi alata, odobrenja i prelaza stanja.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Kako se evaluira?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Testovi ishoda + putanje + robusnosti.\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Šta se menja nakon ažuriranja modela\u002Fruntime-a?\u003C\u002Ftd>\u003Ctd class=\"border border-gray-300 px-4 py-2\">Regresioni paket za izbor alata, dozvole, zaustavljanje i oporavak.\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Ch2 id=\"section-113\">Rubni slučajevi i ograničenja\u003C\u002Fh2>\n\u003Cp>Neki sistemi su „agentni“ samo u uskom smislu rutiranja: model izabere jednog specijalistu ili alat, a zatim je ostatak toka rada deterministički. To i dalje može biti korisno, ali ne treba ga opisivati kao ekvivalent dugotrajnom autonomnom agentu.\u003C\u002Fp>\n\u003Cp>Domeni sa visokim posledicama mogu namerno da ograniče autonomiju agenta. AI sistem može da pregleda dokaze, pripremi preporuke i popuni strukturirane obrasce, dok čovek ostaje jedini akter koji sme da izvrši konačnu transakciju.\u003C\u002Fp>\n\u003Cp>Neka okruženja su dobro prilagođena agentima jer je povratna informacija objektivna. Agenti za kodiranje mogu da pokreću testove; infrastrukturni agenti mogu da pregledaju metrike; agenti za podatke mogu da validiraju rezultate upita. Otvoreni domeni sa slabom povratnom informacijom zahtevaju oprezniju evaluaciju.\u003C\u002Fp>\n\u003Cp>Agent može da radi potpuno lokalno, potpuno kroz upravljane cloud usluge ili u hibridnoj arhitekturi. Agentno ponašanje opisuje tok kontrole, a ne lokaciju hostovanja.\u003C\u002Fp>\n\u003Cp>Termin „rezonovanje“ ne treba koristiti kao dokaz da je unutrašnji proces agenta ispravan. Produkcijsko osiguranje treba da se oslanja na uočljive ulaze, radnje, izlaze, stanje i evaluaciju, a ne na neproverljive tvrdnje o skrivenom rezonovanju.\u003C\u002Fp>\n\u003Ch2 id=\"section-119\">Šta bi promenilo ovaj odgovor?\u003C\u002Fh2>\n\u003Cp>API-ji dobavljača i agentni okviri će nastaviti da se razvijaju, ali granica arhitekture je stabilna: model predlaže odluke, runtime upravlja petljom, alati se povezuju sa okruženjem, dozvole ograničavaju radnje, a spoljna zapažanja određuju šta se zaista dogodilo.\u003C\u002Fp>\n\u003Cp>Kako modeli postaju pouzdaniji, sistemi mogu bezbedno da delegiraju duže horizonte ili složenije ponašanje oporavka. Kako se verifikacija i autorizacija u runtime-u poboljšavaju, neki koraci odobrenja mogu postati automatizovani. To su promene u nivou autonomije, a ne promene u osnovnim slojevima odgovornosti.\u003C\u002Fp>\n\u003Cp>Preporučena arhitektura se takođe menja u zavisnosti od posledica. Istraživački agent koji samo čita javne izvore može da toleriše drugačije kontrole od agenta koji piše produkcijsku konfiguraciju ili prenosi novac.\u003C\u002Fp>\n\u003Ch2 id=\"section-123\">Povezano kanonsko znanje\u003C\u002Fh2>\n\u003Cp>Agentna AI se nalazi iznad nekoliko preduslovnih slojeva: inženjering konteksta određuje šta model vidi; arhitektura izvora istine određuje koje informacije su autoritativne; pretraga obezbeđuje spoljne dokaze; runtime arhitektura određuje šta može da se izvrši.\u003C\u002Fp>\n\u003Cp>Nizvodni čvorovi uključuju pozivanje alata, MCP, A2A, identitet agenta, dozvole, revizibilnost, čoveka u petlji, orkestraciju, memoriju i multi-agent sisteme.\u003C\u002Fp>\n\u003Cp>Članak o protokolarnom steku stoga treba čitati nakon osnovnog koncepta agenta: protokoli standardizuju granice oko agenata; oni ne definišu samo agentno ponašanje.\u003C\u002Fp>\n\u003Ch2 id=\"section-127\">Često postavljana pitanja\u003C\u002Fh2>\n\u003Csection class=\"editorjs-faq my-6 rounded-xl border border-gray-200 p-5 dark:border-gray-700\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">Česta pitanja o agentnoj veštačkoj inteligenciji\u003C\u002Fh3>\u003Cdiv id=\"faq1\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">Šta je agentna veštačka inteligencija?\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">Agentna veštačka inteligencija je AI sistem u kojem model može da teži cilju kroz više koraka birajući akcije ili alate, posmatrajući rezultate, ažurirajući svoje stanje i nastavljajući dok se ne dostigne uslov za zaustavljanje.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq2\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">Koja je razlika između LLM-a i AI agenta?\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">LLM proizvodi izlaze na osnovu ulaza. Agent kombinuje model sa izvršnim okruženjem, alatima, stanjem, dozvolama, upravljanjem kontekstom i iterativnom petljom izvršavanja.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq3\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">Da li pozivanje alata čini sistem agentom?\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">Ne nužno. Jedan odgovor modela uz pomoć alata može biti ograničen i ne-agentan. Agentno ponašanje se pojavljuje kada posmatranja alata pokreću adaptivnu petlju sa više koraka.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq4\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">Koja je razlika između agenta i AI radnog toka?\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">Radni tok obično prati putanju procesa definisanu u kodu aplikacije. Agent ima više kontrole koju pokreće model nad tim koje korake i alate da koristi na osnovu posmatranja između koraka.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq5\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">Da li su agentima potrebna pamćenja?\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">Ne. Dugoročno pamćenje je korisno za trajne informacije između sesija, ali mnogi agenti završavaju ograničene zadatke koristeći samo trenutno stanje zadatka i kontekst.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq6\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">Da li AI agentima treba više agenata?\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">Ne. Jedan agent je često jednostavniji. Sistemi sa više agenata su opravdani kada specijalizacija, izolacija alata, izolacija politika ili granice vlasništva materijalno poboljšavaju sistem.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq7\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">Može li agent da uključuje čoveka u petlji?\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">Da. Agent može autonomno da obavlja analizu i pripremu niskog rizika, dok izvršno okruženje pravi pauzu za ljudsko odobrenje pre akcija sa posledicama.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq8\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">Da li je MCP agentni okvir?\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">Ne. MCP je protokol interoperabilnosti za izlaganje alata, resursa i upita. Izvršno okruženje agenta može da koristi MCP, ali i dalje zahteva sopstvenu petlju, stanje, autorizaciju i evaluaciju.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003Cdiv id=\"faq9\" class=\"border-t border-gray-200 py-4 first:border-t-0 dark:border-gray-700\">\u003Ch4 class=\"font-semibold text-gray-900 dark:text-gray-100\">Kako znate da je agent zaista završio zadatak?\u003C\u002Fh4>\u003Cdiv class=\"mt-2 text-gray-600 dark:text-gray-300\">Gde je moguće, proverite uspeh kroz eksterno stanje, testove, artefakte ili merodavne sistemske zapise, umesto da verujete modelovoj sopstvenoj izjavi o završetku.\u003C\u002Fdiv>\u003C\u002Fdiv>\u003C\u002Fsection>\n\u003Ch2 id=\"section-129\">Pojmovnik\u003C\u002Fh2>\n\u003Csection class=\"editorjs-glossary my-6 rounded-xl border border-gray-200 dark:border-gray-700 p-5\">\u003Ch3 class=\"mb-3 text-lg font-semibold\">Ključni pojmovi agentne veštačke inteligencije\u003C\u002Fh3>\u003Cdl>\u003Cdiv id=\"agentic-ai\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Agentna veštačka inteligencija\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Ponašanje AI sistema u kojem model dinamički usmerava izvršavanje u više koraka koristeći alate, posmatranja i stanje ka cilju.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"ai-agent\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">AI agent\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Sistem sa modelom u središtu, sa izvršnim okruženjem, alatima, stanjem i petljom izvršavanja koja može da teži zadatku kroz više koraka.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"agent-loop\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Agentna petlja\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Ponovljeni ciklus odluke modela, izvršavanja alata\u002Fakcije, posmatranja i ažurirane odluke modela do zaustavljanja.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"runtime-harness\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Izvršno okruženje \u002F okvir\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Sloj izvršavanja koji upravlja petljom modela, alatima, stanjem, odobrenjima, kontekstom, greškama i uslovima za zaustavljanje.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"tool\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Alat\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Sposobnost izložena modelu za čitanje informacija, računanje, delegiranje ili menjanje eksternog stanja.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"observation\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Posmatranje\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Informacija vraćena iz alata ili okruženja i dostavljena kasnijem koraku agenta.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"agent-state\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Stanje agenta\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Trajna informacija o zadatku ili izvršavanju koja postoji izvan jednog izlaza modela i može da preživi kroz korake ili pauze.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"autonomy-boundary\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Granica autonomije\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Eksplicitna granica koja definiše kojim odlukama i akcijama model može dinamički da upravlja.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"human-in-the-loop\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Čovek u petlji\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Obrazac kontrole u kojem je ljudski pregled, unos ili odobrenje potrebno u odabranim tačkama AI vođenog procesa.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"trajectory\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Trajektorija\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Niz relevantnih stanja, odluka, poziva alata, akcija i posmatranja između zahteva zadatka i konačnog ishoda.\u003C\u002Fdd>\u003C\u002Fdiv>\u003Cdiv id=\"idempotency\" class=\"border-t border-gray-200 dark:border-gray-700 py-3 first:border-t-0\">\u003Cdt class=\"font-semibold text-gray-900 dark:text-gray-100\">Idempotentnost\u003C\u002Fdt>\u003Cdd class=\"mt-1 text-gray-600 dark:text-gray-300\">Svojstvo koje omogućava da se operacija ponovi bez nenamernog višestrukog primenjivanja iste sporedne posledice.\u003C\u002Fdd>\u003C\u002Fdiv>\u003C\u002Fdl>\u003C\u002Fsection>\n\u003Ch2 id=\"section-131\">Zaključak\u003C\u002Fh2>\n\u003Cp>Agentna veštačka inteligencija nije samo pametniji model ili chatbot sa više alata. To je sistemska arhitektura u kojoj model učestvuje u iterativnoj kontrolnoj petlji: odluči, deluj, posmatraj, ažuriraj i nastavi.\u003C\u002Fp>\n\u003Cp>Model obezbeđuje fleksibilno donošenje odluka, ali okolno izvršno okruženje mora da poseduje stvarnost izvršavanja: dozvole, pristup alatima, stanje, odobrenja, ponovne pokušaje, budžete, uslove za zaustavljanje, praćenje i verifikaciju.\u003C\u002Fp>\n\u003Cp>Najkorisniji princip dizajna je stoga: prepustite taktički izbor modelu samo unutar eksplicitnih tehničkih i poslovnih granica. Agentna sposobnost postaje proizvodna sposobnost samo kada autonomija, ovlašćenje i dokaz ostanu razdvojivi.\u003C\u002Fp>\n\u003Ch2 id=\"section-135\">Primarni izvori i aktuelne smernice\u003C\u002Fh2>\n\u003Cp>Izvori u nastavku podržavaju aktuelne arhitektonske razlike oko agenata, radnih tokova, petlji, alata, orkestracije, bezbednosti i evaluacije. Sekcije projekta su originalni dokazi implementacije i eksplicitno su ograničene na ono što repozitorijumi pokazuju.\u003C\u002Fp>\n\u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">OpenAI — Agenti\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">Aktuelne smernice za programere koje definišu izbore izvršnog okruženja za rad u više koraka, alate, stanje, orkestraciju i izvršavanje agenata.\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\u002Fdefine-agents\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">OpenAI — Definicije agenata\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">Aktuelna dokumentacija koja opisuje agenta kao model plus instrukcije i opciono ponašanje izvršnog okruženja uključujući alate, zaštitne mere, MCP servere i predaje.\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\u002Frunning-agents\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">OpenAI — Pokretanje agenata\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">Aktuelna dokumentacija agentne petlje: poziv modela, izvršavanje alata ili predaja, nastavak i konačna tačka zaustavljanja.\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\u002Forchestration\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">OpenAI — Orkestracija i predaje\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">Aktuelne smernice o predajama, agentima kao alatima i kada specijalizovani agenti dodaju korisne granice vlasništva ili sposobnosti.\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagent-builder-safety\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">OpenAI — Bezbednost u izgradnji agenata\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">Aktuelne bezbednosne smernice koje pokrivaju odobrenja alata, ubacivanje upita, zaštitne mere i evaluaciju zasnovanu na tragu.\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Anthropic — Izgradnja efikasnih agenata\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">Inženjerske smernice koje razlikuju unapred definisane radne tokove od agenata kojima upravlja model i opisuju petlje povratne informacije iz okruženja zasnovane na alatima.\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Feffective-context-engineering-for-ai-agents\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Anthropic — Efikasno inženjerstvo konteksta za AI agente\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">Praktično određenje agenata kao LLM-ova koji autonomno koriste alate u petlji, uz dinamičko upravljanje kontekstom baš na vreme.\u003C\u002Fp>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fdemystifying-evals-for-ai-agents\" target=\"_blank\" rel=\"noopener noreferrer\" class=\"editorjs-link-tool block border border-gray-200 dark:border-gray-700 rounded-lg p-4 transition text-gray-900 dark:text-gray-100 hover:border-primary-500 hover:bg-primary-50 dark:hover:bg-gray-900 hover:text-gray-900 dark:hover:text-gray-100\">\u003Cstrong class=\"block font-semibold\">Anthropic — Razjašnjavanje evaluacija za AI agente\u003C\u002Fstrong>\u003Cp class=\"text-sm text-gray-600 dark:text-gray-400\">Smernice iz 2026. o evaluaciji agenata sa više krugova koji pozivaju alate, menjaju stanje i prilagođavaju se međurezultatima.\u003C\u002Fp>\u003C\u002Fa>",{"time":212,"blocks":213,"version":1538},1791481653024,[214,220,228,235,242,250,255,260,265,270,275,280,285,290,319,324,329,334,339,371,376,381,414,419,424,468,473,478,483,490,495,500,505,510,515,544,549,554,559,564,569,574,579,584,589,594,599,605,610,615,620,625,630,635,644,649,654,659,664,689,694,699,704,709,714,749,754,759,764,769,774,779,784,789,794,799,804,812,817,822,827,832,840,845,885,890,895,900,905,953,958,987,992,997,1002,1007,1012,1017,1022,1027,1032,1037,1042,1068,1074,1079,1123,1128,1166,1171,1213,1218,1264,1269,1274,1279,1284,1289,1294,1299,1304,1309,1314,1319,1324,1329,1334,1339,1381,1386,1435,1440,1445,1450,1455,1460,1465,1475,1484,1493,1502,1511,1520,1529],{"id":215,"data":216,"type":218,"tunes":219},"intro",{"text":217},"Agentna AI je AI sistem u kojem model može da teži cilju kroz više koraka tako što odlučuje šta će sledeće uraditi, koristeći alate ili druge sposobnosti, posmatrajući rezultate, ažurirajući svoje radno stanje i nastavljajući dok ne dostigne uslov za zaustavljanje. Sam model nije agent. Upotrebljiv agent takođe zahteva izvršno okruženje ili okvir koji upravlja kontekstom, izvršavanjem alata, stanjem, dozvolama, odobrenjima, greškama i petljom između odluka i posmatranja.","paragraph",{},{"id":221,"data":222,"type":226,"tunes":227},"direct",{"body":223,"title":224,"variant":225},"Normalan poziv modela je obično \u003Cstrong>ulaz → model → izlaz\u003C\u002Fstrong>. Agentni sistem je bliži \u003Cstrong>cilj → odluka → alat\u002Fakcija → posmatranje → ažurirana odluka → … → rezultat\u003C\u002Fstrong>.\u003Cbr>\u003Cbr>Ključna razlika nije u tome da li aplikacija koristi LLM ili pozivanje funkcija. Već u tome da li sistem daje modelu značajnu kontrolu nad sledećim korakom višekoraknog procesa, dok izvršno okruženje ograničava šta model zapravo sme da radi.","Direktan odgovor","info","callout",{},{"id":229,"data":230,"type":226,"tunes":234},"boundary",{"body":231,"title":232,"variant":233},"Model može da zna kako da pozove alat. Izvršno okruženje može da izloži taj alat. Nijedna od tih činjenica ne znači da je trenutni korisnik ili agent ovlašćen da izvrši osnovnu poslovnu akciju. \u003Cstrong>Sposobnost alata, dozvola za alat i poslovno ovlašćenje su odvojeni slojevi.\u003C\u002Fstrong>","Sposobnost nije ovlašćenje","warning",{},{"id":236,"data":237,"type":226,"tunes":241},"current",{"body":238,"title":239,"variant":240},"Terminologija agenata još uvek varira među dobavljačima i istraživačkim zajednicama. OpenAI trenutno definiše izvršna okruženja agenata oko višekoraknog rada, alata, stanja i orkestracije. Anthropic-ova praktična razlika ostaje korisna: radni tokovi prate unapred definisane putanje koda, dok agenti dinamički usmeravaju sopstveni proces i korišćenje alata. Ovaj članak stoga tretira „agentnu AI“ kao arhitektonski spektar, a ne kao jednu standardizovanu kategoriju proizvoda.","Napomena o aktuelnim izvorima — 8. oktobar 2026.","note",{},{"id":243,"data":244,"type":248,"tunes":249},"toc",{"title":245,"maxLevel":246,"minLevel":247},"Sadržaj",3,2,"tableOfContents",{},{"id":251,"data":252,"type":42,"tunes":254},"h-meaning",{"text":253,"level":247},"Šta agentna AI zaista znači",{},{"id":256,"data":257,"type":218,"tunes":259},"p-meaning-1",{"text":258},"Važan pomak od obične generativne AI ka agentnoj AI je kontrola nad procesom. Normalan asistent može da odgovori na pitanje koristeći kontekst koji dobije. Agent može da odluči da odgovaranje zahteva dodatne korake: pregledanje datoteke, pretragu repozitorijuma, upit ka API-ju, traženje pojašnjenja, pokretanje testa, ažuriranje tiketa, delegiranje podzadatka ili ponovni pokušaj nakon neuspele akcije.",{},{"id":261,"data":262,"type":218,"tunes":264},"p-meaning-2",{"text":263},"To ne zahteva neograničenu autonomiju. Agent može da radi unutar uskog sandbox-a, pod strogim dozvolama, uz odobrenje potrebno pre svake konsekventne akcije. Sistem je i dalje agentan ako model dinamički bira među dozvoljenim sledećim koracima.",{},{"id":266,"data":267,"type":218,"tunes":269},"p-meaning-3",{"text":268},"Arhitektura je stoga važnija od etikete. „Agent“ treba da opisuje ponašanje sistema: iterativno odlučivanje vođeno modelom nad alatima, stanjem i povratnim informacijama — a ne samo chatbot sa većim promptom.",{},{"id":271,"data":272,"type":42,"tunes":274},"h-simple",{"text":273,"level":247},"Najjednostavniji primer",{},{"id":276,"data":277,"type":218,"tunes":279},"p-simple-1",{"text":278},"Pretpostavimo da programer pita AI sistem: „Pronađi zašto test suite ne prolazi i ispravi grešku.“ Jedan poziv modela mogao bi samo da predloži verovatne uzroke na osnovu teksta koji je dobio.",{},{"id":281,"data":282,"type":218,"tunes":284},"p-simple-2",{"text":283},"Agentni sistem za kodiranje može da pregleda repozitorijum, pronađe test koji ne prolazi, pročita relevantne datoteke, predloži izmenu, uredi kod, pokrene test, posmatra neuspeh, revidira implementaciju i ponovo pokrene test.",{},{"id":286,"data":287,"type":218,"tunes":289},"p-simple-3",{"text":288},"Agentni deo nije samo to što postoje shell i datotečni alati. Već to što model može da koristi povratne informacije iz okruženja da izabere sledeći korak umesto da prati jedan potpuno unapred definisan niz.",{},{"id":291,"data":292,"type":317,"tunes":318},"simple-loop",{"steps":293,"title":315,"orientation":316},[294,297,300,303,306,309,312],{"label":295,"description":296},"1. Primanje cilja","Korisnik ili nadređeni sistem definiše cilj i relevantna ograničenja.",{"label":298,"description":299},"2. Izgradnja trenutnog konteksta","Izvršno okruženje obezbeđuje instrukcije, stanje, istoriju, memoriju, alate i trenutne dokaze.",{"label":301,"description":302},"3. Model odlučuje o sledećem koraku","Model može da odgovori, pozove alat, zatraži informacije, delegira ili se zaustavi.",{"label":304,"description":305},"4. Izvršno okruženje validira zahtev","Dozvole, šeme, odobrenja i politika određuju da li predložena akcija sme da se izvrši.",{"label":307,"description":308},"5. Izvršavanje alata ili akcije","Spoljašnje okruženje se menja ili vraća nove informacije.",{"label":310,"description":311},"6. Posmatranje rezultata","Izvršno okruženje vraća strukturisani izlaz alata, greške ili promene stanja u sledeći korak modela.",{"label":313,"description":314},"7. Nastavak ili zaustavljanje","Petlja se ponavlja dok ne postigne uspeh, odbijanje, eskalaciju, ograničenje budžeta, istek vremena ili drugi uslov za zaustavljanje.","Osnovna petlja agenta","auto","processFlow",{},{"id":320,"data":321,"type":42,"tunes":323},"h-stops",{"text":322,"level":247},"Gde se jednostavan primer zaustavlja",{},{"id":325,"data":326,"type":218,"tunes":328},"p-stops-1",{"text":327},"Nije svaki višekorakni AI sistem jednako agentan. Radni tok može da koristi nekoliko LLM poziva i alata, dok je svaki korak unapred određen u kodu. Drugi sistem može da dozvoli modelu da odluči koji alat će pozvati, kojim redosledom, koliko puta i kada će se zaustaviti.",{},{"id":330,"data":331,"type":218,"tunes":333},"p-stops-2",{"text":332},"Oba mogu biti korisna. Razlika je u tome gde se nalazi kontrola. Unapred definisani radni tokovi stavljaju više kontrole u aplikacijski kod. Agenti premeštaju više taktičkih odluka o procesu u petlju model\u002Fizvršno okruženje.",{},{"id":335,"data":336,"type":42,"tunes":338},"h-workflow",{"text":337,"level":247},"Agent naspram radnog toka",{},{"id":340,"data":341,"type":369,"tunes":370},"workflow-comparison",{"rows":342,"title":360,"layout":361,"columns":362},[343,348,352,356],{"id":344,"label":345,"values":346},"path","Put procesa",[347,347],"",{"id":349,"label":350,"values":351},"tools","Sekvenca alata",[347,347],{"id":353,"label":354,"values":355},"strength","Snaga",[347,347],{"id":357,"label":358,"values":359},"risk","Rizik",[347,347],"Predefinisani tok rada i agentska kontrola","table",[363,366],{"id":364,"label":365},"workflow","LLM tok rada",{"id":367,"label":368},"agent","Agent","comparison",{},{"id":372,"data":373,"type":218,"tunes":375},"p-workflow-1",{"text":374},"Anthropic eksplicitno razdvaja ova dva obrasca: tokovi rada orkestriraju modele i alate kroz predefinisane putanje koda, dok agenti dozvoljavaju modelima da dinamički usmeravaju sopstvene procese i korišćenje alata. Ovo nije jedina moguća terminologija, ali predstavlja korisnu arhitektonsku granicu.",{},{"id":377,"data":378,"type":42,"tunes":380},"h-spectrum",{"text":379,"level":247},"Agentsko ponašanje je spektar, a ne binarna oznaka",{},{"id":382,"data":383,"type":361,"tunes":413},"spectrum-table",{"content":384,"stretched":43,"withHeadings":14},[385,389,393,397,401,405,409],[386,387,388],"Nivo","Primer","Ko odlučuje o sledećem koraku?",[390,391,392],"Poziv jednog modela","Sumiraj ovaj dokument","Aplikacija poziva model jednom",[394,395,396],"Odgovor uz pomoć alata","Model može koristiti web pretragu pre odgovaranja","Model bira iz ograničenih alata za jedan odgovor",[398,399,400],"Strukturirani tok rada","Klasifikuj → preuzmi → generiši → validiraj","Tok rada aplikacije određuje faze",[402,403,404],"Adaptivni tok rada","Model može birati između nekoliko grana i ponoviti","Deljena kontrola između aplikacije i modela",[406,407,408],"Agentska petlja","Model više puta bira alate\u002Fradnje na osnovu zapažanja","Model usmerava taktičko izvršavanje unutar ograničenja izvršnog okruženja",[410,411,412],"Dugotrajni agent","Agent pauzira, nastavlja, upravlja artefaktima i nastavlja dalje","Model + trajno izvršno okruženje upravljaju evoluirajućim izvršavanjem",{},{"id":415,"data":416,"type":218,"tunes":418},"p-spectrum-1",{"text":417},"Nazivanje svakog sistema iznad „agentom“ može prikriti važne operativne razlike. Što je jača kontrola modela nad sekvencom, trajanjem i radnjama, to su važniji izolacija izvršnog okruženja, dozvole, praćenje, uslovi zaustavljanja i evaluacija trajektorije.",{},{"id":420,"data":421,"type":42,"tunes":423},"h-anatomy",{"text":422,"level":247},"Minimalna arhitektura agentskog sistema",{},{"id":425,"data":426,"type":361,"tunes":467},"anatomy-table",{"content":427,"stretched":43,"withHeadings":14},[428,431,434,437,440,443,446,449,452,455,458,461,464],[429,430],"Komponenta","Odgovornost",[432,433],"Cilj \u002F zadatak","Definiše šta sistem pokušava da postigne.",[435,436],"Model","Tumači kontekst i odlučuje o sledećoj radnji ili izlazu.",[438,439],"Instrukcije","Definišu ulogu, ograničenja, prioritete i politiku specifičnu za zadatak.",[441,442],"Sastavljač konteksta","Gradi informacije vidljive modelu u svakom koraku.",[444,445],"Katalog alata","Definiše mogućnosti koje model može zahtevati.",[447,448],"Izvršno okruženje \u002F okvir","Pokreće petlju, izvršava alate, upravlja stanjem i rukuje uslovima zaustavljanja.",[450,451],"Sloj autorizacije","Određuje da li je predložena radnja dozvoljena za trenutnog korisnika.",[453,454],"Stanje \u002F sesija","Čuva napredak zadatka kroz različite faze ili korake izvršavanja.",[456,457],"Kanal zapažanja","Vraća rezultate alata i promene okruženja u sledeći korak modela.",[459,460],"Odobrenja \u002F ljudska kontrola","Pauzira radnje sa posledicama kada je potrebna provera.",[462,463],"Praćenje \u002F revizija","Beleži pozive modela, alate, prelaze, odobrenja i neuspehe.",[465,466],"Evaluacija","Meri ishode i trajektorije izvršavanja u odnosu na kriterijume prihvatanja.",{},{"id":469,"data":470,"type":42,"tunes":472},"h-model-agent",{"text":471,"level":247},"Model nije agent",{},{"id":474,"data":475,"type":218,"tunes":477},"p-model-agent-1",{"text":476},"Jezički model proizvodi izlaze na osnovu ulaza. On sam po sebi ne poseduje fajl sistem, ne izvršava shell komandu, ne održava trajno stanje zadatka, ne sprovodi dozvole niti se automatski ponovo poziva.",{},{"id":479,"data":480,"type":218,"tunes":482},"p-model-agent-2",{"text":481},"Te sposobnosti dolaze iz okolnog izvršnog okruženja. Isti model može se ponašati kao jednostavan chat model u jednoj aplikaciji i kao motor za odlučivanje unutar agentske petlje u drugoj.",{},{"id":484,"data":485,"type":226,"tunes":489},"model-agent-rule",{"body":486,"title":487,"variant":488},"\u003Cstrong>Sposobnost modela određuje koje odluke mogu biti predložene. Arhitektura izvršnog okruženja određuje šta se zapravo može dogoditi.\u003C\u002Fstrong>","Arhitektonsko pravilo","success",{},{"id":491,"data":492,"type":42,"tunes":494},"h-tools",{"text":493,"level":247},"Korišćenje alata je centralno — ali samo korišćenje alata ne čini agenta",{},{"id":496,"data":497,"type":218,"tunes":499},"p-tools-1",{"text":498},"Alati omogućavaju modelu da pribavi informacije i utiče na spoljne sisteme. Primeri uključuju čitanje baze podataka, operacije sa fajlovima, izvršavanje shell komandi, web pretragu, kontrolu pregledača, API pozive, ažuriranje tiketa ili delegiranje specijalizovanim agentima.",{},{"id":501,"data":502,"type":218,"tunes":504},"p-tools-2",{"text":503},"Poziv jednog modela može koristiti jedan alat i dalje ostati ograničen odgovor uz pomoć alata, a ne dugotrajni agent. Agentsko ponašanje se pojavljuje kada zapažanja alata hrane adaptivnu petlju u kojoj model bira šta će sledeće uraditi.",{},{"id":506,"data":507,"type":218,"tunes":509},"p-tools-3",{"text":508},"Dizajn alata je važan jer su alati ugovor između rezonovanja modela i spoljne stvarnosti. Dvosmisleni ili preklapajući alati stvaraju greške u rutiranju; veliki nestrukturirani izlazi zagađuju kontekst; alati sa širokim sporednim efektima povećavaju obim štete.",{},{"id":511,"data":512,"type":42,"tunes":514},"h-capability-permission",{"text":513,"level":247},"Sposobnost alata, dozvola i ovlašćenje su različiti",{},{"id":516,"data":517,"type":361,"tunes":543},"permission-table",{"content":518,"stretched":43,"withHeadings":14},[519,522,525,528,531,534,537,540],[520,521],"Sloj","Pitanje",[523,524],"Sposobnost","Može li ovo izvršno okruženje tehnički izvesti operaciju?",[526,527],"Izloženost alata","Da li je ta sposobnost dostupna ovom agentu?",[529,530],"Dozvola","Sme li ovaj agent\u002Fsesija da je koristi pod trenutnom politikom?",[532,533],"Korisnička autorizacija","Da li je korisnik koji zahteva dozvoljen da izazove ovu operaciju?",[535,536],"Poslovno ovlašćenje","Da li je operacija validna prema pravilima domena, odobrenjima i ograničenjima?",[538,539],"Izvršavanje","Da li se operacija zaista dogodila?",[541,542],"Revizija","Može li sistem dokazati ko je zahtevao, odobrio i izvršio operaciju?",{},{"id":545,"data":546,"type":218,"tunes":548},"p-permission-1",{"text":547},"Ovi slojevi se često spajaju u prototipovima. Model vidi alat za refundaciju i stoga izgleda da može izvršiti refundacije. U produkciji, alat bi ipak trebalo da validira nalog, korisnika, transakciju, iznos, politiku i uslove odobrenja nezavisno od zahteva modela.",{},{"id":550,"data":551,"type":42,"tunes":553},"h-runtime",{"text":552,"level":247},"Izvršno okruženje ili okvir je stvarni sistem za izvršavanje",{},{"id":555,"data":556,"type":218,"tunes":558},"p-runtime-1",{"text":557},"OpenAI-jeva trenutna dokumentacija o agentima eksplicitno pravi razliku između izvršnih okruženja. Različita izvršna okruženja mogu upravljati orkestracijom, stanjem, alatima, sandbox-ovima i izvršavanjem na različitim mestima, dok model ostaje samo jedan deo sistema.",{},{"id":560,"data":561,"type":218,"tunes":563},"p-runtime-2",{"text":562},"Agents SDK opisuje petlju koja više puta poziva trenutni model, pregleda izlaz, izvršava tražene alate ili predaje, i nastavlja dok model ne vrati konačan odgovor ili drugu stvarnu tačku zaustavljanja.",{},{"id":565,"data":566,"type":218,"tunes":568},"p-runtime-3",{"text":567},"To znači da odluke o arhitekturi agenta uključuju gde se izvršava orkestracija, gde se čuva stanje, ko izvršava alate, koji sandbox sadrži neželjene efekte i ko upravlja ponovnim pokušajima, vremenskim ograničenjima i mogućnošću nastavljanja.",{},{"id":570,"data":571,"type":42,"tunes":573},"h-planning",{"text":572,"level":247},"Planiranje je korisno, ali eksplicitni plan nije neophodan",{},{"id":575,"data":576,"type":218,"tunes":578},"p-planning-1",{"text":577},"Agenti se često opisuju kao sistemi koji „planiraju“. U praksi, planiranje može biti eksplicitno ili implicitno. Agent može prvo da proizvede vidljiv višekoračni plan, ili može da bira jednu sledeću akciju po jednu i revidira nakon svakog zapažanja.",{},{"id":580,"data":581,"type":218,"tunes":583},"p-planning-2",{"text":582},"Za veoma neizvesne zadatke, planiranje kratkog horizonta može biti sigurnije jer okruženje može poništiti dugoročni plan. Arhitektonski zahtev je sposobnost izbora i revizije akcija na osnovu cilja, trenutnog stanja i novih dokaza.",{},{"id":585,"data":586,"type":42,"tunes":588},"h-feedback",{"text":587,"level":247},"Povratne informacije iz okruženja čine petlju korisnom",{},{"id":590,"data":591,"type":218,"tunes":593},"p-feedback-1",{"text":592},"Agent postaje operativno značajan kada može da primeti da li je njegova akcija uspela. Izlaz alata, rezultati testova, odgovori API-ja, stanje fajl sistema, stanje pregledača i zapisi aplikacija pružaju spoljne dokaze koje sistem može da koristi za reviziju sledeće odluke.",{},{"id":595,"data":596,"type":218,"tunes":598},"p-feedback-2",{"text":597},"Anthropic-ove smernice za agente naglašavaju ovu povratnu petlju: agenti koriste alate, dobijaju osnovnu istinu iz okruženja, procenjuju napredak i nastavljaju ili traže ljudski unos.",{},{"id":600,"data":601,"type":226,"tunes":604},"feedback-rule",{"body":602,"title":603,"variant":233},"To što agent kaže „zadatak je završen“ ne dokazuje završetak. Gde je moguće, proverite konačno stanje kroz spoljni sistem, test, fajl, zapis transakcije ili drugi vidljivi ishod.","Samoizveštavanje nije dokaz iz okruženja",{},{"id":606,"data":607,"type":42,"tunes":609},"h-state",{"text":608,"level":247},"Stanje agenta nije isto kao kontekst modela",{},{"id":611,"data":612,"type":218,"tunes":614},"p-state-1",{"text":613},"Dugotrajan zadatak može zahtevati stanje koje ne može ili ne bi trebalo da ostane u kontekstu modela: ID-ovi zadataka, kontrolne tačke, artefakti, odobrenja, identifikatori spoljnih objekata, brojači ponovnih pokušaja i status radnog toka.",{},{"id":616,"data":617,"type":218,"tunes":619},"p-state-2",{"text":618},"Izvršno okruženje može da sačuva ovo trajno stanje izvan prozora modela i rekonstruiše kontekst potreban za sledeći korak. To održava kontekst vidljiv modelu fokusiranim dok se održava kontinuitet i mogućnost nastavljanja.",{},{"id":621,"data":622,"type":42,"tunes":624},"h-memory",{"text":623,"level":247},"Memorija je opciona, nije definicija agenta",{},{"id":626,"data":627,"type":218,"tunes":629},"p-memory-1",{"text":628},"Agent može uspešno da radi bez dugoročne memorije ako kompletan zadatak staje unutar jednog ograničenog izvršavanja. Memorija postaje korisna kada informacije moraju da se zadrže kroz sesije, zadatke ili duge horizonte izvršavanja.",{},{"id":631,"data":632,"type":218,"tunes":634},"p-memory-2",{"text":633},"RAG, memorija, stanje i kontekst rešavaju različite probleme. Tretiranje vektorske baze podataka kao „memorije agenta“ ili istorije razgovora kao „mašine stanja“ obično skriva važne granice životnog ciklusa i autoriteta.",{},{"id":636,"data":637,"type":642,"tunes":643},"ref-memory",{"url":638,"title":639,"excerpt":640,"ctaLabel":641},"https:\u002F\u002Fstajic.de\u002Fsr\u002Fblog\u002Fai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context","Memorija AI agenta nije RAG: Kako razdvojiti memoriju, pretragu, stanje i kontekst","Praktična arhitektura koja razdvaja trajnu memoriju, autoritativno stanje aplikacije, pretragu i kontekst koji se dostavlja modelu.","Pročitajte članak o arhitekturi memorije","referralArticle",{},{"id":645,"data":646,"type":42,"tunes":648},"h-context",{"text":647,"level":247},"Inženjering konteksta postaje dinamičan u agentima",{},{"id":650,"data":651,"type":218,"tunes":653},"p-context-1",{"text":652},"Svaki poziv alata može da proizvede novi kontekst. Svaki korak takođe može da učini ranije informacije zastarelim. Snažno izvršno okruženje agenta zato ponovo gradi ili oblikuje kontekst kako izvršavanje napreduje, umesto da iznova prikazuje sve neograničeno.",{},{"id":655,"data":656,"type":218,"tunes":658},"p-context-2",{"text":657},"Definicije alata, stanje zadatka, preuzeti dokazi, zapažanja i memorija svi se takmiče za pažnju modela. Dugotrajni agenti zahtevaju skraćivanje, sažimanje ili učitavanje na vreme kako bi kontekst ostao relevantan za trenutnu odluku.",{},{"id":660,"data":661,"type":42,"tunes":663},"h-sideeffects",{"text":662,"level":247},"Alati za čitanje i alati sa sporednim efektima imaju različit rizik",{},{"id":665,"data":666,"type":369,"tunes":688},"sideeffect-comparison",{"rows":667,"title":680,"layout":361,"columns":681},[668,672,676],{"id":669,"label":670,"values":671},"examples","Primeri",[347,347],{"id":673,"label":674,"values":675},"mainrisk","Glavni rizik",[347,347],{"id":677,"label":678,"values":679},"control","Tipična kontrola",[347,347],"Pristup informacijama naspram spoljne akcije",[682,685],{"id":683,"label":684},"read","Čitanje \u002F posmatranje",{"id":686,"label":687},"write","Pisanje \u002F delovanje",{},{"id":690,"data":691,"type":42,"tunes":693},"h-approval",{"text":692,"level":247},"Čovek u petlji je kontrolni mehanizam, a ne suprotnost agentnoj AI",{},{"id":695,"data":696,"type":218,"tunes":698},"p-approval-1",{"text":697},"Agent ne prestaje da bude agentan zato što čovek odobrava konsekventne korake. Model i dalje može autonomno da pregleda, zaključuje, pretražuje i priprema akciju dok izvršno okruženje zahteva ljudsku potvrdu pre izvršenja.",{},{"id":700,"data":701,"type":218,"tunes":703},"p-approval-2",{"text":702},"OpenAI-ove trenutne smernice za bezbednost agenata eksplicitno preporučuju odobrenja za operacije alata u rizičnijim radnim tokovima. Anthropic takođe naglašava kontrolne tačke i ljudsku procenu tamo gde agenti naiđu na prepreke ili konsekventne odluke.",{},{"id":705,"data":706,"type":218,"tunes":708},"p-approval-3",{"text":707},"Korisno arhitektonsko pitanje nije „čovek ili autonomno?“ već koje odluke mogu da se delegiraju, koje zahtevaju pregled i koje moraju da ostanu determinističke?",{},{"id":710,"data":711,"type":42,"tunes":713},"h-stopping",{"text":712,"level":247},"Agenti zahtevaju eksplicitne uslove zaustavljanja",{},{"id":715,"data":716,"type":361,"tunes":748},"stop-table",{"content":717,"stretched":43,"withHeadings":14},[718,721,724,727,730,733,736,739,742,745],[719,720],"Uslov zaustavljanja","Svrha",[722,723],"Uspešan verifikovan ishod","Završi kada je spoljno ciljno stanje potvrđeno.",[725,726],"Maksimalan broj koraka","Spreči nekontrolisane petlje.",[728,729],"Vremenski budžet","Ograniči izvršavanje po stvarnom vremenu.",[731,732],"Budžet troškova\u002Ftokena","Ograniči potrošnju resursa.",[734,735],"Detektor ponovljene akcije","Zaustavi petlje koje više ne napreduju.",[737,738],"Granica dozvola","Pauziraj ili zaustavi kada sledeća potrebna akcija nije dozvoljena.",[740,741],"Kontrolna tačka ljudskog odobrenja","Sačekaj pre konsekventnog izvršenja.",[743,744],"Neopoziv kvar alata","Eskaliraj umesto beskonačnog ponavljanja.",[746,747],"Prag neizvesnosti","Zatraži pojašnjenje kada zadatak ne može bezbedno da se zaključi.",{},{"id":750,"data":751,"type":42,"tunes":753},"h-errors",{"text":752,"level":247},"Oporavak je deo ponašanja agenta",{},{"id":755,"data":756,"type":218,"tunes":758},"p-errors-1",{"text":757},"Agenti rade u okruženjima koja otkazuju: API-ji isteknu, datoteke se promene, akreditivi isteknu, web stranice se premeste i alati vraćaju neispravan izlaz. Koristan agentni sistem zato zahteva ponašanje oporavka, a ne samo petlju alata za idealan scenario.",{},{"id":760,"data":761,"type":218,"tunes":763},"p-errors-2",{"text":762},"Oporavak može da uključi ponovni pokušaj sa ograničenjima, izbor drugog alata, ponovno čitanje trenutnog stanja, pitanje korisniku, vraćanje delimične akcije ili eskalaciju ka čoveku.",{},{"id":765,"data":766,"type":218,"tunes":768},"p-errors-3",{"text":767},"Ponovni pokušaji takođe zahtevaju svest o idempotentnosti. Ponavljanje čitanja je obično niskog rizika; ponavljanje plaćanja ili slanja poruke može da stvori duplikate sporednih efekata.",{},{"id":770,"data":771,"type":42,"tunes":773},"h-single-multi",{"text":772,"level":247},"Agentna AI ne zahteva više agenata",{},{"id":775,"data":776,"type":218,"tunes":778},"p-multi-1",{"text":777},"Jedan agent sa jasnim skupom alata je često jednostavniji i lakši za evaluaciju od multi-agent arhitekture. Više agenata je korisno kada specijalizacija materijalno poboljšava izolaciju alata, izolaciju politika, jasnoću upita, vlasništvo ili čitljivost traga.",{},{"id":780,"data":781,"type":218,"tunes":783},"p-multi-2",{"text":782},"OpenAI-ove trenutne smernice za orkestraciju eksplicitno preporučuju početak sa jednim agentom gde je to moguće i dodavanje specijalista samo kada se ugovor ili granica vlasništva materijalno promene.",{},{"id":785,"data":786,"type":218,"tunes":788},"p-multi-3",{"text":787},"Multi-agentni sistemi dodaju nove probleme: kvalitet delegiranja, duplirani kontekst, konfliktno stanje, semantiku predaje, identitet, troškove i distribuirano rukovanje greškama.",{},{"id":790,"data":791,"type":42,"tunes":793},"h-protocols",{"text":792,"level":247},"Agent protokoli su slojevi interoperabilnosti, a ne sam agent",{},{"id":795,"data":796,"type":218,"tunes":798},"p-protocols-1",{"text":797},"Protokoli kao što su MCP i A2A mogu učiniti arhitekturu agenta interoperabilnom, ali sami po sebi ne stvaraju petlju agenta. MCP može izložiti alate i resurse. A2A može povezati nezavisno implementirane agente. Aplikaciji je i dalje potrebno izvršno okruženje, autorizacija, stanje, evaluacija i domenska logika.",{},{"id":800,"data":801,"type":218,"tunes":803},"p-protocols-2",{"text":802},"Zato sposobnost protokola mora ostati odvojena od poslovnog ovlašćenja. Otkrivanje alata putem MCP-a ne dokazuje da trenutni principal sme da ga koristi. Prijem zadatka putem A2A ne dokazuje da udaljeni agent sme da izvrši svaku traženu radnju.",{},{"id":805,"data":806,"type":642,"tunes":811},"ref-protocols",{"url":807,"title":808,"excerpt":809,"ctaLabel":810},"https:\u002F\u002Fstajic.de\u002Fde\u002Fblog\u002Fmcp-vs-a2a-vs-ucp-vs-ap2-vs-a2ui-the-agent-protocol-stack-explained","MCP vs A2A vs UCP vs AP2 vs A2UI: Objašnjen stek agent protokola","Mapa odgovornosti protokola koja pokazuje zašto pristup alatima, saradnja agenata, trgovina, ovlašćenje plaćanja i UI vođen agentom pripadaju različitim granicama interoperabilnosti.","Pročitajte stek agent protokola",{},{"id":813,"data":814,"type":42,"tunes":816},"h-reliability",{"text":815,"level":247},"Trajektorija je deo pouzdanosti agenta",{},{"id":818,"data":819,"type":218,"tunes":821},"p-rel-1",{"text":820},"Konačan odgovor je nedovoljan dokaz za agentni sistem jer agent može doći do ispravnog rezultata preko nesigurne ili nevalidne putanje. Može koristiti neovlašćeni alat, preskočiti obaveznu proveru, ponoviti sporedni efekat, osloniti se na zastarelo stanje ili slučajno uspeti.",{},{"id":823,"data":824,"type":218,"tunes":826},"p-rel-2",{"text":825},"Evaluacija zato zahteva tragove izvršavanja: odluke, pozive alata, odobrenja, opservacije, promene stanja i konačan ishod. Trenutne OpenAI smernice za bezbednost preporučuju ocenjivače tragova i evaluacije; Anthropic-ove smernice za evaluaciju agenata iz 2026. takođe tretiraju višekratne trajektorije alata kao evaluacione objekte prvog reda.",{},{"id":828,"data":829,"type":218,"tunes":831},"p-rel-3",{"text":830},"Jače pitanje pouzdanosti je: Da li je agent postigao prihvatljiv ishod kroz prihvatljivu, obnovljivu i revizibilnu trajektoriju?",{},{"id":833,"data":834,"type":642,"tunes":839},"ref-reliability",{"url":835,"title":836,"excerpt":837,"ctaLabel":838},"https:\u002F\u002Fstajic.de\u002Fsr\u002Fblog\u002Fai-agent-reliability-why-the-final-answer-is-not-enough","Pouzdanost AI agenata: Zašto konačan odgovor nije dovoljan","Zašto proizvodna evaluacija mora da ispituje trajektorije, korišćenje alata, prelaze stanja i obnovljivost, a ne samo konačne odgovore.","Pročitajte članak o pouzdanosti",{},{"id":841,"data":842,"type":42,"tunes":844},"h-security",{"text":843,"level":247},"Agentni sistemi povećavaju bezbednosnu površinu",{},{"id":846,"data":847,"type":361,"tunes":884},"security-table",{"content":848,"stretched":43,"withHeadings":14},[849,852,856,860,864,868,872,876,880],[358,850,851],"Zašto ga agenti pojačavaju","Arhitektonski odgovor",[853,854,855],"Ubacivanje upita","Nepouzdan sadržaj može uticati na buduće odluke o alatima","Odvojite instrukcije od podataka; ograničite alate; sanitizujte ili strukturirajte spoljni unos gde je moguće",[857,858,859],"Prekomerna ovlašćenja","Greške u rezonovanju mogu postati stvarni sporedni efekti","Najmanja privilegija, ograničeni akreditivi, politika po alatu i odobrenja",[861,862,863],"Izlaganje akreditiva","Alatima mogu biti potrebne moćne tajne","Držite tajne van konteksta modela; posredujte pristup kroz pouzdano izvršno okruženje",[865,866,867],"Zbunjeni posrednik","Agent može delovati sa ovlašćenjem širim od korisnika koji zahteva","Vežite izvršavanje za identitet korisnika\u002Fservisa i ponovo autorizujte posledične radnje",[869,870,871],"Petlje bez kontrole","Model više puta poziva alate bez napretka","Budžeti za korake, vreme i troškove plus detekcija petlji",[873,874,875],"Pomeranje stanja","Okruženje se menja nakon što agent formira plan","Ponovo pročitajte autoritativno stanje pre posledičnih radnji",[877,878,879],"Indirektno ubacivanje","Sadržaj alata\u002Fveba\u002Fdokumenta sadrži instrukcije usmerene na model","Tretirajte spoljni sadržaj kao nepouzdane podatke, a ne kao instrukcijsko ovlašćenje",[881,882,883],"Praznina u reviziji","Konačan rezultat ne može pokazati šta je izvršeno","Pratite pozive alata, odobrenja, identitete i promene stanja",{},{"id":886,"data":887,"type":42,"tunes":889},"h-observability",{"text":888,"level":247},"Opservabilnost agenta mora pratiti petlju",{},{"id":891,"data":892,"type":218,"tunes":894},"p-obs-1",{"text":893},"Tradicionalna opservabilnost servisa beleži zahteve, latenciju i greške. Opservabilnost agenta zahteva dodatni model izvršavanja: koji agent je bio aktivan, koja verzija modela je donela odluku, koji kontekst je bio dostupan, koji alat je izabran, koji argumenti su poslati, koji rezultat je vraćen i zašto je izvršavanje zaustavljeno.",{},{"id":896,"data":897,"type":218,"tunes":899},"p-obs-2",{"text":898},"Za osetljive sisteme, sami tragovi zahtevaju kontrolu pristupa i politiku zadržavanja jer upiti, izlazi alata i artefakti mogu sadržati poverljive podatke.",{},{"id":901,"data":902,"type":42,"tunes":904},"h-eval",{"text":903,"level":247},"Kako evaluirati agentni sistem",{},{"id":906,"data":907,"type":361,"tunes":952},"eval-table",{"content":908,"stretched":43,"withHeadings":14},[909,912,916,920,924,928,932,936,940,944,948],[910,521,911],"Dimenzija","Primer dokaza",[913,914,915],"Uspeh zadatka","Da li se traženi ishod dogodio?","Spoljno stanje, testovi, poslovni ishod",[917,918,919],"Kvalitet trajektorije","Da li su koraci bili prihvatljivi?","Trag alata\u002Fradnji",[921,922,923],"Izbor alata","Da li je agent izabrao odgovarajuće sposobnosti?","Očekivani vs stvarni pozivi alata",[925,926,927],"Poštovanje dozvola","Da li je ostao unutar dozvoljenog ovlašćenja?","Dnevnici autorizacije i testovi odbijenih radnji",[929,930,931],"Rukovanje stanjem","Da li je koristio trenutno autoritativno stanje?","Provere svežine i testovi promene stanja",[933,934,935],"Oporavak","Da li je ispravno odgovorio na greške?","Ubačeni scenariji tajmauta\u002Fgreške",[937,938,939],"Ponašanje zaustavljanja","Da li se zaustavio na pravom mestu?","Brojevi koraka, detekcija petlji, dokaz konačnog stanja",[941,942,943],"Ljudska eskalacija","Da li je pitao kada je pregled bio potreban?","Tragovi odobrenja\u002Feskalacije",[945,946,947],"Trošak\u002Flatencija","Da li je autonomija vredela operativnog troška?","Tokeni, pozivi alata, trajanje",[949,950,951],"Robusnost","Da li preživljava realistične varijacije okruženja?","Ponovljeni i adversarni pokušaji",{},{"id":954,"data":955,"type":42,"tunes":957},"h-use",{"text":956,"level":247},"Kada je agent prikladan",{},{"id":959,"data":960,"type":361,"tunes":986},"use-table",{"content":961,"stretched":43,"withHeadings":14},[962,965,968,971,974,977,980,983],[963,964],"Koristite agenta kada","Preferirajte radni tok ili jednostavan poziv kada",[966,967],"Broj ili redosled koraka ne može biti pouzdano poznat unapred","Sekvenca je stabilna i deterministička",[969,970],"Sistem mora da pregleda okruženje i prilagodi se","Jedan korak preuzimanja i generisanja je dovoljan",[972,973],"Nekoliko alata može biti korisno u zavisnosti od međurezultata","Jedan poznati API poziv rešava zadatak",[975,976],"Zadatak ima koristi od iterativne verifikacije ili ispravke","Odgovor može biti proizveden direktno iz dostavljenog konteksta",[978,979],"Neuspesi zahtevaju fleksibilno ponašanje oporavka","Grane neuspeha su jednostavne i mogu se eksplicitno kodirati",[981,982],"Ljudski pregled može biti umetnut na smislenim kontrolnim tačkama","Svaki korak je visokorizičan i mora se ionako ručno kontrolisati",[984,985],"Očekivana vrednost opravdava dodatnu latenciju, trošak i složenost","Predvidivost i niska cena su važniji od fleksibilnosti",{},{"id":988,"data":989,"type":218,"tunes":991},"p-use-1",{"text":990},"Snažan podrazumevani pristup je započeti sa najjednostavnijim rešenjem koje funkcioniše i povećavati agentsku složenost samo kada fleksibilnost proizvodi merljivu vrednost. Agenti razmenjuju predvidivost, latenciju i trošak za adaptivno izvršavanje.",{},{"id":993,"data":994,"type":42,"tunes":996},"h-implementation",{"text":995,"level":247},"Dokazi iz originalne implementacije",{},{"id":998,"data":999,"type":42,"tunes":1001},"h-client",{"text":1000,"level":246},"Aaasaasa AI Client: model, izvršno okruženje i dozvole su odvojeni",{},{"id":1003,"data":1004,"type":218,"tunes":1006},"p-client-1",{"text":1005},"Aaasaasa AI Client eksplicitno razdvaja agenta\u002Fklijenta, provajdera, model, lokaciju izvršnog okruženja i dozvole. Njegova arhitektonska dokumentacija tretira dozvole kao centralnu politiku alata\u002Fradnog prostora, a ne kao osobinu modela.",{},{"id":1008,"data":1009,"type":218,"tunes":1011},"p-client-2",{"text":1010},"Ista aplikacija može da izloži Direct Chat bez alata za fajl sistem ili ljusku, dok Codex izvršno okruženje radi pod izabranim radnim prostorom i profilom dozvola. Ovo demonstrira ključnu granicu agentske arhitekture: promena površine izvršnog okruženja\u002Falata menja šta sistem može da uradi čak i kada pristup modelu ostaje dostupan.",{},{"id":1013,"data":1014,"type":218,"tunes":1016},"p-client-3",{"text":1015},"Repozitorijum takođe razlikuje lokalno Codex izvršno okruženje od lokacije modela: lokalno izvršno okruženje može pozvati model u oblaku. Ovo sprečava uobičajenu grešku izjednačavanja „agent se izvršava lokalno“ sa „inferencija je lokalna“.",{},{"id":1018,"data":1019,"type":218,"tunes":1021},"p-client-4",{"text":1020},"Implementacija onemogućava ugrađene putanje izvršavanja čija semantika odobravanja ne zadovoljava zahtevani model dozvola. Ovo podržava princip da sposobnost agenta ne treba da zaobiđe autorizaciju izvršnog okruženja samo zato što osnovni okvir može da izvršava alate.",{},{"id":1023,"data":1024,"type":42,"tunes":1026},"h-sot-agent",{"text":1025,"level":246},"Source of Truth Research Engine: ograničene agentske faze istraživanja",{},{"id":1028,"data":1029,"type":218,"tunes":1031},"p-sot-1",{"text":1030},"Source of Truth Research Engine koristi ograničen istraživački tok: otkrij → pribavi → izvuci → verifikuj → protivreči → sintetiši. Istraživački poslovi mogu se izvršavati kroz AI izvršno okruženje dok dokazi, izvori, tvrdnje i protivrečnosti ostaju u eksternom trajnom skladištu.",{},{"id":1033,"data":1034,"type":218,"tunes":1036},"p-sot-2",{"text":1035},"Ovo je namerno kontrolisanije od neograničenog autonomnog istraživačkog agenta. Faze pružaju zaštitne ograde oko toga koja vrsta posla treba sledeće da se obavi, dok i dalje dozvoljavaju istraživanje vođeno modelom unutar svakog ograničenog zadatka.",{},{"id":1038,"data":1039,"type":218,"tunes":1041},"p-sot-3",{"text":1040},"Ta razlika je koristan dokaz za dizajn agenata: autonomija može biti smeštena unutar strukturisanog okvira isporuke, umesto da se primenjuje uniformno na ceo proces.",{},{"id":1043,"data":1044,"type":361,"tunes":1067},"impl-table",{"content":1045,"stretched":43,"withHeadings":14},[1046,1049,1052,1055,1058,1061,1064],[1047,1048],"Implementirani obrazac","Lekcija agentske arhitekture",[1050,1051],"Direct Chat nema OS alate","Model može postojati bez agentske sposobnosti izvršavanja.",[1053,1054],"Codex izvršno okruženje ima profil dozvola radnog prostora","Ovlašćenje alata pripada politici izvršnog okruženja, a ne sposobnosti modela.",[1056,1057],"Provajder\u002Fmodel\u002Fizvršno okruženje su odvojeni koncepti","Lokacija agentskog okvira i lokacija inferencije su nezavisne odluke.",[1059,1060],"Broker dozvola za izvršna okruženja sa alatima","Izlaganje sposobnosti može biti centralizovano i upravljano.",[1062,1063],"Ograničene faze istraživanja","Autonomija može da radi unutar eksplicitnih granica procesa.",[1065,1066],"Trajne tvrdnje\u002Fdokazi izvan konteksta modela","Stanje agenta i dokazi ne moraju da žive samo u istoriji razgovora.",{},{"id":1069,"data":1070,"type":226,"tunes":1073},"impl-boundary",{"body":1071,"title":1072,"variant":240},"Ovi projekti demonstriraju konkretne obrasce agenta\u002Fizvršnog okruženja, dozvola i ograničenog istraživanja. Nisu predstavljeni kao dokaz masovnog komercijalnog uvođenja autonomnih agenata.","Granica dokaza",{},{"id":1075,"data":1076,"type":42,"tunes":1078},"h-failures",{"text":1077,"level":247},"Uobičajeni načini neuspeha agentske AI",{},{"id":1080,"data":1081,"type":361,"tunes":1122},"failure-table",{"content":1082,"stretched":43,"withHeadings":14},[1083,1086,1089,1092,1095,1098,1101,1104,1107,1110,1113,1116,1119],[1084,1085],"Način neuspeha","Šta je zapravo zakazalo",[1087,1088],"„Agent“ je samo chatbot sa alatima navedenim u upitu","Ne postoji pouzdana petlja izvršnog okruženja ili arhitektura izvršavanja alata",[1090,1091],"Podrška za alate se tretira kao dozvola","Granice sposobnosti i autorizacije su srušene",[1093,1094],"Agent veruje sopstvenoj izjavi o završetku","Ishod nije verifikovan u odnosu na eksterno stanje",[1096,1097],"Svaki zadatak postaje multi-agent","Složenost raste bez stvarne granice vlasništva ili specijalizacije",[1099,1100],"Istorija razgovora se koristi kao trajno stanje","Mogućnost nastavka i autoritativno stanje postaju krhki",[1102,1103],"Agent slepo ponavlja sporedne efekte","Duplikati poruka, plaćanja ili promena stanja postaju mogući",[1105,1106],"Nema ograničenja koraka\u002Ftroškova","Agent može da se vrti u nedogled ili troši nekontrolisane resurse",[1108,1109],"Izlaz alata se veruje kao instrukcija","Indirektna injekcija upita može preusmeriti ponašanje",[1111,1112],"Tačan konačni odgovor je jedina evaluacija","Nesigurne ili nevažeće putanje ostaju nevidljive",[1114,1115],"Nadogradnja modela se tretira kao transparentna","Izbor alata, planiranje i ponašanje zaustavljanja mogu se promeniti",[1117,1118],"Jedan širok alat izlaže mnoge privilegovane operacije","Radijus eksplozije se povećava i nameru je teže validirati",[1120,1121],"Ljudsko odobrenje postoji, ali recenzent nema kontekst","Odobrenje postaje ceremonijalno, a ne efektivno",{},{"id":1124,"data":1125,"type":42,"tunes":1127},"h-misconceptions",{"text":1126,"level":247},"Uobičajene zablude",{},{"id":1129,"data":1130,"type":361,"tunes":1165},"misconceptions-table",{"content":1131,"stretched":43,"withHeadings":14},[1132,1135,1138,1141,1144,1147,1150,1153,1156,1159,1162],[1133,1134],"Zabluda","Ispravka",[1136,1137],"„LLM je agent.“","Model je komponenta odlučivanja; agent je okolni sistem koji upravlja alatima, stanjem i iteracijom.",[1139,1140],"„Pozivanje alata automatski znači agentska AI.“","Jedan ograničen poziv alata možda ne uključuje adaptivnu agentsku petlju sa više koraka.",[1142,1143],"„Agenti moraju biti potpuno autonomni.“","Agentski sistemi mogu zahtevati odobrenja i raditi unutar uskih granica dozvola.",[1145,1146],"„Agenti moraju imati dugoročnu memoriju.“","Memorija je opciona; mnogi korisni agenti završavaju ograničene zadatke bez memorije između sesija.",[1148,1149],"„Agenti moraju prvo napraviti pisani plan.“","Planiranje može biti eksplicitno ili implicitno i može se odvijati korak po korak.",[1151,1152],"„Multi-agent je napredniji od single-agent.“","Složeniji je; koristite ga samo kada specijalizacija ili granice vlasništva to opravdavaju.",[1154,1155],"„MCP stvara agenta.“","MCP izlaže alate\u002Fresurse; izvršno okruženje i dalje zahteva agentsku petlju i model autorizacije.",[1157,1158],"„Lokalno izvršno okruženje znači da je model lokalni.“","Lokacija izvršnog okruženja i lokacija inferencije\u002Fprovajdera su odvojene.",[1160,1161],"„Ako je konačni rezultat tačan, agent je radio ispravno.“","Nesigurna ili neovlašćena putanja i dalje može proizvesti tačan rezultat.",[1163,1164],"„Ljudsko odobrenje uklanja autonomiju.“","Odobrenje može ograničiti izabrane akcije dok ostatak procesa ostaje vođen modelom.",{},{"id":1167,"data":1168,"type":42,"tunes":1170},"h-design",{"text":1169,"level":247},"Praktičan redosled dizajna agenta",{},{"id":1172,"data":1173,"type":317,"tunes":1212},"design-flow",{"steps":1174,"title":1211,"orientation":316},[1175,1178,1181,1184,1187,1190,1193,1196,1199,1202,1205,1208],{"label":1176,"description":1177},"1. Definišite ishod","Navedite koji spoljni rezultat ili artefakt dokazuje uspeh zadatka.",{"label":1179,"description":1180},"2. Odlučite da li je agent zaista potreban","Preferirajte jednostavan poziv ili deterministički tok rada kada je put predvidiv.",{"label":1182,"description":1183},"3. Identifikujte stanje i izvor istine","Definišite koji sistemi poseduju trenutne činjenice, napredak zadatka i poslovno stanje.",{"label":1185,"description":1186},"4. Definišite površinu alata","Izložite najmanji skup jasnih sposobnosti potrebnih za zadatak.",{"label":1188,"description":1189},"5. Povežite identitet i dozvole","Razdvojite korisničko ovlašćenje, dozvole agenta\u002Fruntime-a i sposobnosti alata.",{"label":1191,"description":1192},"6. Izaberite granice autonomije","Navedite šta model može dinamički da odlučuje, a šta ostaje deterministički.",{"label":1194,"description":1195},"7. Dodajte kontrolne tačke za odobrenje","Zahtevajte pregled pre posledičnih ili nepovratnih radnji gde je to prikladno.",{"label":1197,"description":1198},"8. Definišite zaustavljanje i oporavak","Postavite dokaz uspeha, budžete, tajmaute, ponovne pokušaje, eskalaciju i kontrolu petlji.",{"label":1200,"description":1201},"9. Dizajnirajte upravljanje kontekstom\u002Fstanjem","Držite trenutno stanje, memoriju, zapažanja alata i trajne artefakte u ispravnim slojevima.",{"label":1203,"description":1204},"10. Pratite putanju","Zabeležite dovoljno strukture izvršavanja za otklanjanje grešaka i reviziju odluka modela\u002Falata.",{"label":1206,"description":1207},"11. Procenite realistične neuspehe","Testirajte zastarelo stanje, greške alata, ubacivanje upita, dvosmislene zahteve i promenjena okruženja.",{"label":1209,"description":1210},"12. Proširite autonomiju samo na osnovu dokaza","Povećajte dozvole ili horizont izvršavanja kada evaluacija pokaže da korist opravdava rizik.","Dizajnirajte agenta od ovlašćenja ka spolja",{},{"id":1214,"data":1215,"type":42,"tunes":1217},"h-checklist",{"text":1216,"level":247},"Kontrolna lista arhitekture agentne AI",{},{"id":1219,"data":1220,"type":361,"tunes":1263},"checklist-table",{"content":1221,"stretched":43,"withHeadings":14},[1222,1224,1227,1230,1233,1236,1239,1242,1245,1248,1251,1254,1257,1260],[521,1223],"Očekivani dokaz",[1225,1226],"Šta dokazuje uspeh?","Spoljni ishod, artefakt, test ili autoritativno stanje.",[1228,1229],"Zašto je agent potreban?","Put zaista zavisi od posrednih zapažanja.",[1231,1232],"Koje odluke pokreće model?","Eksplicitna granica autonomije.",[1234,1235],"Koji alati postoje?","Mali, dokumentovan, nedvosmislen skup sposobnosti.",[1237,1238],"Ko može da koristi svaki alat?","Politika autorizacije svesna identiteta i konteksta.",[1240,1241],"Koje radnje zahtevaju odobrenje?","Pravila pregleda zasnovana na posledicama.",[1243,1244],"Gde se čuva stanje zadatka?","Stanje u vlasništvu aplikacije, odvojeno od prolaznog konteksta modela.",[1246,1247],"Kako se agent oporavlja?","Ponašanje ponovnog pokušaja, ponovnog čitanja, vraćanja, pojašnjenja i eskalacije.",[1249,1250],"Kako se zaustavlja?","Verifikovani završetak plus ograničenja koraka\u002Fvremena\u002Ftroškova.",[1252,1253],"Kako su zaštićeni sporedni efekti?","Validacija, idempotentnost, najmanja privilegija i potvrda.",[1255,1256],"Može li se izvršavanje rekonstruisati?","Tragovi alata, odobrenja i prelaza stanja.",[1258,1259],"Kako se evaluira?","Testovi ishoda + putanje + robusnosti.",[1261,1262],"Šta se menja nakon ažuriranja modela\u002Fruntime-a?","Regresioni paket za izbor alata, dozvole, zaustavljanje i oporavak.",{},{"id":1265,"data":1266,"type":42,"tunes":1268},"h-edge",{"text":1267,"level":247},"Rubni slučajevi i ograničenja",{},{"id":1270,"data":1271,"type":218,"tunes":1273},"p-edge-1",{"text":1272},"Neki sistemi su „agentni“ samo u uskom smislu rutiranja: model izabere jednog specijalistu ili alat, a zatim je ostatak toka rada deterministički. To i dalje može biti korisno, ali ne treba ga opisivati kao ekvivalent dugotrajnom autonomnom agentu.",{},{"id":1275,"data":1276,"type":218,"tunes":1278},"p-edge-2",{"text":1277},"Domeni sa visokim posledicama mogu namerno da ograniče autonomiju agenta. AI sistem može da pregleda dokaze, pripremi preporuke i popuni strukturirane obrasce, dok čovek ostaje jedini akter koji sme da izvrši konačnu transakciju.",{},{"id":1280,"data":1281,"type":218,"tunes":1283},"p-edge-3",{"text":1282},"Neka okruženja su dobro prilagođena agentima jer je povratna informacija objektivna. Agenti za kodiranje mogu da pokreću testove; infrastrukturni agenti mogu da pregledaju metrike; agenti za podatke mogu da validiraju rezultate upita. Otvoreni domeni sa slabom povratnom informacijom zahtevaju oprezniju evaluaciju.",{},{"id":1285,"data":1286,"type":218,"tunes":1288},"p-edge-4",{"text":1287},"Agent može da radi potpuno lokalno, potpuno kroz upravljane cloud usluge ili u hibridnoj arhitekturi. Agentno ponašanje opisuje tok kontrole, a ne lokaciju hostovanja.",{},{"id":1290,"data":1291,"type":218,"tunes":1293},"p-edge-5",{"text":1292},"Termin „rezonovanje“ ne treba koristiti kao dokaz da je unutrašnji proces agenta ispravan. Produkcijsko osiguranje treba da se oslanja na uočljive ulaze, radnje, izlaze, stanje i evaluaciju, a ne na neproverljive tvrdnje o skrivenom rezonovanju.",{},{"id":1295,"data":1296,"type":42,"tunes":1298},"h-change",{"text":1297,"level":247},"Šta bi promenilo ovaj odgovor?",{},{"id":1300,"data":1301,"type":218,"tunes":1303},"p-change-1",{"text":1302},"API-ji dobavljača i agentni okviri će nastaviti da se razvijaju, ali granica arhitekture je stabilna: model predlaže odluke, runtime upravlja petljom, alati se povezuju sa okruženjem, dozvole ograničavaju radnje, a spoljna zapažanja određuju šta se zaista dogodilo.",{},{"id":1305,"data":1306,"type":218,"tunes":1308},"p-change-2",{"text":1307},"Kako modeli postaju pouzdaniji, sistemi mogu bezbedno da delegiraju duže horizonte ili složenije ponašanje oporavka. Kako se verifikacija i autorizacija u runtime-u poboljšavaju, neki koraci odobrenja mogu postati automatizovani. To su promene u nivou autonomije, a ne promene u osnovnim slojevima odgovornosti.",{},{"id":1310,"data":1311,"type":218,"tunes":1313},"p-change-3",{"text":1312},"Preporučena arhitektura se takođe menja u zavisnosti od posledica. Istraživački agent koji samo čita javne izvore može da toleriše drugačije kontrole od agenta koji piše produkcijsku konfiguraciju ili prenosi novac.",{},{"id":1315,"data":1316,"type":42,"tunes":1318},"h-related",{"text":1317,"level":247},"Povezano kanonsko znanje",{},{"id":1320,"data":1321,"type":218,"tunes":1323},"p-related-1",{"text":1322},"Agentna AI se nalazi iznad nekoliko preduslovnih slojeva: inženjering konteksta određuje šta model vidi; arhitektura izvora istine određuje koje informacije su autoritativne; pretraga obezbeđuje spoljne dokaze; runtime arhitektura određuje šta može da se izvrši.",{},{"id":1325,"data":1326,"type":218,"tunes":1328},"p-related-2",{"text":1327},"Nizvodni čvorovi uključuju pozivanje alata, MCP, A2A, identitet agenta, dozvole, revizibilnost, čoveka u petlji, orkestraciju, memoriju i multi-agent sisteme.",{},{"id":1330,"data":1331,"type":218,"tunes":1333},"p-related-3",{"text":1332},"Članak o protokolarnom steku stoga treba čitati nakon osnovnog koncepta agenta: protokoli standardizuju granice oko agenata; oni ne definišu samo agentno ponašanje.",{},{"id":1335,"data":1336,"type":42,"tunes":1338},"h-faq",{"text":1337,"level":247},"Često postavljana pitanja",{},{"id":1340,"data":1341,"type":1340,"tunes":1380},"faq",{"items":1342,"title":1379},[1343,1347,1351,1355,1359,1363,1367,1371,1375],{"id":1344,"answer":1345,"question":1346},"faq1","Agentna veštačka inteligencija je AI sistem u kojem model može da teži cilju kroz više koraka birajući akcije ili alate, posmatrajući rezultate, ažurirajući svoje stanje i nastavljajući dok se ne dostigne uslov za zaustavljanje.","Šta je agentna veštačka inteligencija?",{"id":1348,"answer":1349,"question":1350},"faq2","LLM proizvodi izlaze na osnovu ulaza. Agent kombinuje model sa izvršnim okruženjem, alatima, stanjem, dozvolama, upravljanjem kontekstom i iterativnom petljom izvršavanja.","Koja je razlika između LLM-a i AI agenta?",{"id":1352,"answer":1353,"question":1354},"faq3","Ne nužno. Jedan odgovor modela uz pomoć alata može biti ograničen i ne-agentan. Agentno ponašanje se pojavljuje kada posmatranja alata pokreću adaptivnu petlju sa više koraka.","Da li pozivanje alata čini sistem agentom?",{"id":1356,"answer":1357,"question":1358},"faq4","Radni tok obično prati putanju procesa definisanu u kodu aplikacije. Agent ima više kontrole koju pokreće model nad tim koje korake i alate da koristi na osnovu posmatranja između koraka.","Koja je razlika između agenta i AI radnog toka?",{"id":1360,"answer":1361,"question":1362},"faq5","Ne. Dugoročno pamćenje je korisno za trajne informacije između sesija, ali mnogi agenti završavaju ograničene zadatke koristeći samo trenutno stanje zadatka i kontekst.","Da li su agentima potrebna pamćenja?",{"id":1364,"answer":1365,"question":1366},"faq6","Ne. Jedan agent je često jednostavniji. Sistemi sa više agenata su opravdani kada specijalizacija, izolacija alata, izolacija politika ili granice vlasništva materijalno poboljšavaju sistem.","Da li AI agentima treba više agenata?",{"id":1368,"answer":1369,"question":1370},"faq7","Da. Agent može autonomno da obavlja analizu i pripremu niskog rizika, dok izvršno okruženje pravi pauzu za ljudsko odobrenje pre akcija sa posledicama.","Može li agent da uključuje čoveka u petlji?",{"id":1372,"answer":1373,"question":1374},"faq8","Ne. MCP je protokol interoperabilnosti za izlaganje alata, resursa i upita. Izvršno okruženje agenta može da koristi MCP, ali i dalje zahteva sopstvenu petlju, stanje, autorizaciju i evaluaciju.","Da li je MCP agentni okvir?",{"id":1376,"answer":1377,"question":1378},"faq9","Gde je moguće, proverite uspeh kroz eksterno stanje, testove, artefakte ili merodavne sistemske zapise, umesto da verujete modelovoj sopstvenoj izjavi o završetku.","Kako znate da je agent zaista završio zadatak?","Česta pitanja o agentnoj veštačkoj inteligenciji",{},{"id":1382,"data":1383,"type":42,"tunes":1385},"h-glossary",{"text":1384,"level":247},"Pojmovnik",{},{"id":1387,"data":1388,"type":1387,"tunes":1434},"glossary",{"title":1389,"entries":1390},"Ključni pojmovi agentne veštačke inteligencije",[1391,1395,1399,1403,1406,1410,1414,1418,1422,1426,1430],{"term":1392,"anchor":1393,"definition":1394},"Agentna veštačka inteligencija","agentic-ai","Ponašanje AI sistema u kojem model dinamički usmerava izvršavanje u više koraka koristeći alate, posmatranja i stanje ka cilju.",{"term":1396,"anchor":1397,"definition":1398},"AI agent","ai-agent","Sistem sa modelom u središtu, sa izvršnim okruženjem, alatima, stanjem i petljom izvršavanja koja može da teži zadatku kroz više koraka.",{"term":1400,"anchor":1401,"definition":1402},"Agentna petlja","agent-loop","Ponovljeni ciklus odluke modela, izvršavanja alata\u002Fakcije, posmatranja i ažurirane odluke modela do zaustavljanja.",{"term":447,"anchor":1404,"definition":1405},"runtime-harness","Sloj izvršavanja koji upravlja petljom modela, alatima, stanjem, odobrenjima, kontekstom, greškama i uslovima za zaustavljanje.",{"term":1407,"anchor":1408,"definition":1409},"Alat","tool","Sposobnost izložena modelu za čitanje informacija, računanje, delegiranje ili menjanje eksternog stanja.",{"term":1411,"anchor":1412,"definition":1413},"Posmatranje","observation","Informacija vraćena iz alata ili okruženja i dostavljena kasnijem koraku agenta.",{"term":1415,"anchor":1416,"definition":1417},"Stanje agenta","agent-state","Trajna informacija o zadatku ili izvršavanju koja postoji izvan jednog izlaza modela i može da preživi kroz korake ili pauze.",{"term":1419,"anchor":1420,"definition":1421},"Granica autonomije","autonomy-boundary","Eksplicitna granica koja definiše kojim odlukama i akcijama model može dinamički da upravlja.",{"term":1423,"anchor":1424,"definition":1425},"Čovek u petlji","human-in-the-loop","Obrazac kontrole u kojem je ljudski pregled, unos ili odobrenje potrebno u odabranim tačkama AI vođenog procesa.",{"term":1427,"anchor":1428,"definition":1429},"Trajektorija","trajectory","Niz relevantnih stanja, odluka, poziva alata, akcija i posmatranja između zahteva zadatka i konačnog ishoda.",{"term":1431,"anchor":1432,"definition":1433},"Idempotentnost","idempotency","Svojstvo koje omogućava da se operacija ponovi bez nenamernog višestrukog primenjivanja iste sporedne posledice.",{},{"id":1436,"data":1437,"type":42,"tunes":1439},"h-conclusion",{"text":1438,"level":247},"Zaključak",{},{"id":1441,"data":1442,"type":218,"tunes":1444},"p-conclusion-1",{"text":1443},"Agentna veštačka inteligencija nije samo pametniji model ili chatbot sa više alata. To je sistemska arhitektura u kojoj model učestvuje u iterativnoj kontrolnoj petlji: odluči, deluj, posmatraj, ažuriraj i nastavi.",{},{"id":1446,"data":1447,"type":218,"tunes":1449},"p-conclusion-2",{"text":1448},"Model obezbeđuje fleksibilno donošenje odluka, ali okolno izvršno okruženje mora da poseduje stvarnost izvršavanja: dozvole, pristup alatima, stanje, odobrenja, ponovne pokušaje, budžete, uslove za zaustavljanje, praćenje i verifikaciju.",{},{"id":1451,"data":1452,"type":218,"tunes":1454},"p-conclusion-3",{"text":1453},"Najkorisniji princip dizajna je stoga: prepustite taktički izbor modelu samo unutar eksplicitnih tehničkih i poslovnih granica. Agentna sposobnost postaje proizvodna sposobnost samo kada autonomija, ovlašćenje i dokaz ostanu razdvojivi.",{},{"id":1456,"data":1457,"type":42,"tunes":1459},"h-sources",{"text":1458,"level":247},"Primarni izvori i aktuelne smernice",{},{"id":1461,"data":1462,"type":218,"tunes":1464},"p-sources-note",{"text":1463},"Izvori u nastavku podržavaju aktuelne arhitektonske razlike oko agenata, radnih tokova, petlji, alata, orkestracije, bezbednosti i evaluacije. Sekcije projekta su originalni dokazi implementacije i eksplicitno su ograničene na ono što repozitorijumi pokazuju.",{},{"id":1466,"data":1467,"type":1473,"tunes":1474},"src-openai-agents",{"link":1468,"meta":1469},"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents",{"image":1470,"title":1471,"description":1472},{"url":347},"OpenAI — Agenti","Aktuelne smernice za programere koje definišu izbore izvršnog okruženja za rad u više koraka, alate, stanje, orkestraciju i izvršavanje agenata.","linkTool",{},{"id":1476,"data":1477,"type":1473,"tunes":1483},"src-openai-definitions",{"link":1478,"meta":1479},"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\u002Fdefine-agents",{"image":1480,"title":1481,"description":1482},{"url":347},"OpenAI — Definicije agenata","Aktuelna dokumentacija koja opisuje agenta kao model plus instrukcije i opciono ponašanje izvršnog okruženja uključujući alate, zaštitne mere, MCP servere i predaje.",{},{"id":1485,"data":1486,"type":1473,"tunes":1492},"src-openai-running",{"link":1487,"meta":1488},"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\u002Frunning-agents",{"image":1489,"title":1490,"description":1491},{"url":347},"OpenAI — Pokretanje agenata","Aktuelna dokumentacija agentne petlje: poziv modela, izvršavanje alata ili predaja, nastavak i konačna tačka zaustavljanja.",{},{"id":1494,"data":1495,"type":1473,"tunes":1501},"src-openai-orchestration",{"link":1496,"meta":1497},"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\u002Forchestration",{"image":1498,"title":1499,"description":1500},{"url":347},"OpenAI — Orkestracija i predaje","Aktuelne smernice o predajama, agentima kao alatima i kada specijalizovani agenti dodaju korisne granice vlasništva ili sposobnosti.",{},{"id":1503,"data":1504,"type":1473,"tunes":1510},"src-openai-safety",{"link":1505,"meta":1506},"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagent-builder-safety",{"image":1507,"title":1508,"description":1509},{"url":347},"OpenAI — Bezbednost u izgradnji agenata","Aktuelne bezbednosne smernice koje pokrivaju odobrenja alata, ubacivanje upita, zaštitne mere i evaluaciju zasnovanu na tragu.",{},{"id":1512,"data":1513,"type":1473,"tunes":1519},"src-anthropic-agents",{"link":1514,"meta":1515},"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents",{"image":1516,"title":1517,"description":1518},{"url":347},"Anthropic — Izgradnja efikasnih agenata","Inženjerske smernice koje razlikuju unapred definisane radne tokove od agenata kojima upravlja model i opisuju petlje povratne informacije iz okruženja zasnovane na alatima.",{},{"id":1521,"data":1522,"type":1473,"tunes":1528},"src-anthropic-context",{"link":1523,"meta":1524},"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Feffective-context-engineering-for-ai-agents",{"image":1525,"title":1526,"description":1527},{"url":347},"Anthropic — Efikasno inženjerstvo konteksta za AI agente","Praktično određenje agenata kao LLM-ova koji autonomno koriste alate u petlji, uz dinamičko upravljanje kontekstom baš na vreme.",{},{"id":1530,"data":1531,"type":1473,"tunes":1537},"src-anthropic-evals",{"link":1532,"meta":1533},"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fdemystifying-evals-for-ai-agents",{"image":1534,"title":1535,"description":1536},{"url":347},"Anthropic — Razjašnjavanje evaluacija za AI agente","Smernice iz 2026. o evaluaciji agenata sa više krugova koji pozivaju alate, menjaju stanje i prilagođavaju se međurezultatima.",{},"2.31","Agentna AI koristi modele unutar višekoračnih izvršnih petlji gde mogu da biraju alate, posmatraju rezultate, ažuriraju stanje i prilagode svoju sledeću akciju unutar eksplicitnih granica izvršavanja i dozvola.","\u002Fuploads\u002F2026\u002F10\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act-1791481499084-wnji2a.webp","agentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act-1791481499084-wnji2a","PUBLISHED","2026-10-08T11:43:00.000Z","2026-10-08T17:43:28.373Z","2026-10-08T19:19:47.140Z",{"en":1547,"de":1548,"sr":1549,"es":1550,"fr":1551,"it":1552,"ru":1553,"zh":1554},"\u002Fblog\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act","\u002Fde\u002Fblog\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act","\u002Fsr\u002Fblog\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act","\u002Fes\u002Fblog\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act","\u002Ffr\u002Fblog\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act","\u002Fit\u002Fblog\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act","\u002Fru\u002Fblog\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act","\u002Fzh\u002Fblog\u002Fagentic-ai-explained-when-an-ai-system-can-plan-use-tools-and-act",[1556,1560,1564],{"id":1557,"name":1558,"slug":1559},84,"Politike i granice podataka","policy-and-data",{"id":1561,"name":1562,"slug":1563},57,"Granice podataka","data-boundaries",{"id":1565,"name":1566,"slug":1567},59,"Upravljanje i audit","governance",{"id":1569,"login":1570,"email":1571,"displayName":1572},"20","rooth8233","aleksandar@stajic.de","Aleksandar Stajić",[1574,2703],{"lang":1575,"title":1576,"content":1577,"contentJson":1578,"excerpt":2702},"en","Agentic AI Explained: When an AI System Can Plan, Use Tools and Act","{\"time\":1791487185746,\"blocks\":[{\"id\":\"intro\",\"type\":\"paragraph\",\"data\":{\"text\":\"Agentic AI is an AI system in which a model can pursue a goal across multiple steps by deciding what to do next, using tools or other capabilities, observing the results, updating its working state and continuing until it reaches a stopping condition. The model alone is not the agent. A usable agent also needs a runtime or harness that manages context, tool execution, state, permissions, approvals, errors and the loop between decisions and observations.\"},\"tunes\":{}},{\"id\":\"direct\",\"type\":\"callout\",\"data\":{\"variant\":\"info\",\"title\":\"Direct answer\",\"body\":\"A normal model call is usually \u003Cstrong>input → model → output\u003C\u002Fstrong>. An agentic system is closer to \u003Cstrong>goal → decision → tool\u002Faction → observation → updated decision → … → result\u003C\u002Fstrong>.\u003Cbr>\u003Cbr>The critical distinction is not whether an application uses an LLM or function calling. It is whether the system gives the model meaningful control over the next step of a multi-step process while a runtime constrains what the model is actually allowed to do.\"},\"tunes\":{}},{\"id\":\"boundary\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"Capability is not authority\",\"body\":\"A model may know how to call a tool. A runtime may expose that tool. Neither fact means the current user or agent is authorized to execute the underlying business action. \u003Cstrong>Tool capability, tool permission and business authority are separate layers.\u003C\u002Fstrong>\"},\"tunes\":{}},{\"id\":\"current\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"Current-source note — 8 October 2026\",\"body\":\"Agent terminology still varies across vendors and research communities. OpenAI currently defines agent runtimes around multi-step work, tools, state and orchestration. Anthropic's practical distinction remains useful: workflows follow predefined code paths, while agents dynamically direct their own process and tool use. This article therefore treats “agentic AI” as an architectural spectrum rather than one standardized product category.\"},\"tunes\":{}},{\"id\":\"toc\",\"type\":\"tableOfContents\",\"data\":{\"title\":\"Contents\",\"minLevel\":2,\"maxLevel\":3},\"tunes\":{}},{\"id\":\"h-meaning\",\"type\":\"header\",\"data\":{\"text\":\"What agentic AI really means\",\"level\":2},\"tunes\":{}},{\"id\":\"p-meaning-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"The important shift from ordinary generative AI to agentic AI is control over process. A normal assistant can answer a question using the context it receives. An agent can decide that answering requires additional steps: inspect a file, search a repository, query an API, ask for clarification, run a test, update a ticket, delegate a subtask or retry after a failed action.\"},\"tunes\":{}},{\"id\":\"p-meaning-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"This does not require unlimited autonomy. An agent can operate inside a narrow sandbox, under strict permissions, with approval required before every consequential action. The system is still agentic if the model dynamically chooses among permitted next steps.\"},\"tunes\":{}},{\"id\":\"p-meaning-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The architecture therefore matters more than the label. “Agent” should describe a system behavior: iterative model-driven decision making over tools, state and feedback — not merely a chatbot with a larger prompt.\"},\"tunes\":{}},{\"id\":\"h-simple\",\"type\":\"header\",\"data\":{\"text\":\"The simplest example\",\"level\":2},\"tunes\":{}},{\"id\":\"p-simple-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Suppose a developer asks an AI system: “Find why the test suite fails and fix the bug.” A single model call could only suggest likely causes from the text it was given.\"},\"tunes\":{}},{\"id\":\"p-simple-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"An agentic coding system can inspect the repository, search for the failing test, read relevant files, propose a change, edit the code, run the test, observe the failure, revise the implementation and run the test again.\"},\"tunes\":{}},{\"id\":\"p-simple-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The agentic part is not simply that shell and file tools exist. It is that the model can use environmental feedback to choose the next step instead of following one completely predefined sequence.\"},\"tunes\":{}},{\"id\":\"simple-loop\",\"type\":\"processFlow\",\"data\":{\"title\":\"The basic agent loop\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Receive a goal\",\"description\":\"The user or upstream system defines the objective and relevant constraints.\"},{\"label\":\"2. Build current context\",\"description\":\"The runtime supplies instructions, state, history, memory, tools and current evidence.\"},{\"label\":\"3. Model decides next step\",\"description\":\"The model may answer, call a tool, request information, delegate or stop.\"},{\"label\":\"4. Runtime validates the request\",\"description\":\"Permissions, schemas, approvals and policy determine whether the proposed action may execute.\"},{\"label\":\"5. Execute tool or action\",\"description\":\"The external environment changes or returns new information.\"},{\"label\":\"6. Observe the result\",\"description\":\"The runtime feeds structured tool output, errors or state changes back into the next model step.\"},{\"label\":\"7. Continue or stop\",\"description\":\"The loop repeats until success, refusal, escalation, budget limit, timeout or another stopping condition.\"}]},\"tunes\":{}},{\"id\":\"h-stops\",\"type\":\"header\",\"data\":{\"text\":\"Where the simple example stops\",\"level\":2},\"tunes\":{}},{\"id\":\"p-stops-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Not every multi-step AI system is equally agentic. A workflow may use several LLM calls and tools while every step is predetermined in code. Another system may let the model decide which tool to call, in which order, how many times and when to stop.\"},\"tunes\":{}},{\"id\":\"p-stops-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Both can be useful. The difference is where control lives. Predefined workflows put more control in application code. Agents move more tactical process decisions into the model\u002Fruntime loop.\"},\"tunes\":{}},{\"id\":\"h-workflow\",\"type\":\"header\",\"data\":{\"text\":\"Agent vs workflow\",\"level\":2},\"tunes\":{}},{\"id\":\"workflow-comparison\",\"type\":\"comparison\",\"data\":{\"title\":\"Predefined workflow and agentic control\",\"layout\":\"table\",\"columns\":[{\"id\":\"workflow\",\"label\":\"LLM workflow\"},{\"id\":\"agent\",\"label\":\"Agent\"}],\"rows\":[{\"id\":\"path\",\"label\":\"Process path\",\"values\":[\"\",\"\"]},{\"id\":\"tools\",\"label\":\"Tool sequence\",\"values\":[\"\",\"\"]},{\"id\":\"strength\",\"label\":\"Strength\",\"values\":[\"\",\"\"]},{\"id\":\"risk\",\"label\":\"Risk\",\"values\":[\"\",\"\"]}]},\"tunes\":{}},{\"id\":\"p-workflow-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Anthropic explicitly separates these two patterns: workflows orchestrate models and tools through predefined code paths, while agents let models dynamically direct their own processes and tool usage. This is not the only possible terminology, but it is a useful architecture boundary.\"},\"tunes\":{}},{\"id\":\"h-spectrum\",\"type\":\"header\",\"data\":{\"text\":\"Agentic behavior is a spectrum, not a binary label\",\"level\":2},\"tunes\":{}},{\"id\":\"spectrum-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Level\",\"Example\",\"Who decides the next step?\"],[\"Single model call\",\"Summarize this document\",\"Application calls model once\"],[\"Tool-assisted response\",\"Model may use web search before answering\",\"Model selects from bounded tools for one response\"],[\"Structured workflow\",\"Classify → retrieve → generate → validate\",\"Application workflow determines stages\"],[\"Adaptive workflow\",\"Model can choose among several branches and retry\",\"Shared control between application and model\"],[\"Agent loop\",\"Model repeatedly chooses tools\u002Factions based on observations\",\"Model directs tactical execution inside runtime constraints\"],[\"Long-running agent\",\"Agent pauses, resumes, manages artifacts and continues\",\"Model + persistent runtime manage evolving execution\"]]},\"tunes\":{}},{\"id\":\"p-spectrum-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Calling every system above an “agent” can obscure important operational differences. The stronger the model's control over sequence, duration and actions, the more important runtime isolation, permissions, tracing, stopping conditions and trajectory evaluation become.\"},\"tunes\":{}},{\"id\":\"h-anatomy\",\"type\":\"header\",\"data\":{\"text\":\"The minimum architecture of an agentic system\",\"level\":2},\"tunes\":{}},{\"id\":\"anatomy-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Component\",\"Responsibility\"],[\"Goal \u002F task\",\"Defines what the system is trying to accomplish.\"],[\"Model\",\"Interprets context and decides the next action or output.\"],[\"Instructions\",\"Define role, constraints, priorities and task-specific policy.\"],[\"Context assembler\",\"Builds the information visible to the model on each step.\"],[\"Tool catalog\",\"Defines capabilities the model may request.\"],[\"Runtime \u002F harness\",\"Runs the loop, executes tools, manages state and handles stopping conditions.\"],[\"Authorization layer\",\"Determines whether a proposed action is permitted for the current principal.\"],[\"State \u002F session\",\"Preserves task progress across turns or execution steps.\"],[\"Observation channel\",\"Returns tool results and environment changes to the next model step.\"],[\"Approvals \u002F human control\",\"Pauses consequential actions where review is required.\"],[\"Tracing \u002F audit\",\"Records model calls, tools, transitions, approvals and failures.\"],[\"Evaluation\",\"Measures outcomes and execution trajectories against acceptance criteria.\"]]},\"tunes\":{}},{\"id\":\"h-model-agent\",\"type\":\"header\",\"data\":{\"text\":\"A model is not an agent\",\"level\":2},\"tunes\":{}},{\"id\":\"p-model-agent-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A language model produces outputs from inputs. It does not by itself own a filesystem, execute a shell command, maintain durable task state, enforce permissions or automatically call itself again.\"},\"tunes\":{}},{\"id\":\"p-model-agent-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Those capabilities come from the surrounding runtime. The same model can behave as a simple chat model in one application and as the decision engine inside an agent loop in another.\"},\"tunes\":{}},{\"id\":\"model-agent-rule\",\"type\":\"callout\",\"data\":{\"variant\":\"success\",\"title\":\"Architecture rule\",\"body\":\"\u003Cstrong>Model capability determines what decisions can be proposed. Runtime architecture determines what can actually happen.\u003C\u002Fstrong>\"},\"tunes\":{}},{\"id\":\"h-tools\",\"type\":\"header\",\"data\":{\"text\":\"Tool use is central — but tool use alone does not make an agent\",\"level\":2},\"tunes\":{}},{\"id\":\"p-tools-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Tools let the model acquire information and affect external systems. Examples include database reads, file operations, shell execution, web search, browser control, API calls, ticket updates or delegated specialist agents.\"},\"tunes\":{}},{\"id\":\"p-tools-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"A single model call can use one tool and still remain a bounded tool-assisted response rather than a long-running agent. Agentic behavior appears when tool observations feed an adaptive loop in which the model chooses what to do next.\"},\"tunes\":{}},{\"id\":\"p-tools-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Tool design matters because tools are the contract between model reasoning and external reality. Ambiguous or overlapping tools create routing errors; large unstructured outputs pollute context; broad side-effect tools increase blast radius.\"},\"tunes\":{}},{\"id\":\"h-capability-permission\",\"type\":\"header\",\"data\":{\"text\":\"Tool capability, permission and authority are different\",\"level\":2},\"tunes\":{}},{\"id\":\"permission-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Layer\",\"Question\"],[\"Capability\",\"Can this runtime technically perform the operation?\"],[\"Tool exposure\",\"Is that capability available to this agent?\"],[\"Permission\",\"May this agent\u002Fsession use it under the current policy?\"],[\"User authorization\",\"Is the requesting principal allowed to cause this operation?\"],[\"Business authority\",\"Is the operation valid under domain rules, approvals and limits?\"],[\"Execution\",\"Did the operation actually occur?\"],[\"Audit\",\"Can the system prove who requested, approved and executed it?\"]]},\"tunes\":{}},{\"id\":\"p-permission-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"These layers are frequently collapsed in prototypes. A model sees a refund tool and therefore appears able to issue refunds. In production, the tool should still validate account, user, transaction, amount, policy and approval conditions independently of the model's request.\"},\"tunes\":{}},{\"id\":\"h-runtime\",\"type\":\"header\",\"data\":{\"text\":\"The runtime or harness is the actual execution system\",\"level\":2},\"tunes\":{}},{\"id\":\"p-runtime-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"OpenAI's current agent documentation makes the runtime distinction explicit. Different runtimes can manage orchestration, state, tools, sandboxes and execution in different places, while the model remains only one part of the system.\"},\"tunes\":{}},{\"id\":\"p-runtime-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The Agents SDK describes a loop that repeatedly calls the current model, inspects the output, executes requested tools or handoffs, and continues until the model returns a final answer or another real stopping point.\"},\"tunes\":{}},{\"id\":\"p-runtime-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"This means agent architecture decisions include where orchestration runs, where state lives, who executes tools, which sandbox contains side effects, and who owns retries, timeouts and resumability.\"},\"tunes\":{}},{\"id\":\"h-planning\",\"type\":\"header\",\"data\":{\"text\":\"Planning is useful, but an explicit plan is not required\",\"level\":2},\"tunes\":{}},{\"id\":\"p-planning-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Agents are often described as systems that “plan.” In practice, planning can be explicit or implicit. An agent may first produce a visible multi-step plan, or it may choose one next action at a time and revise after every observation.\"},\"tunes\":{}},{\"id\":\"p-planning-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"For highly uncertain tasks, short-horizon planning can be safer because the environment can invalidate a long plan. The architectural requirement is the ability to choose and revise actions based on the goal, current state and new evidence.\"},\"tunes\":{}},{\"id\":\"h-feedback\",\"type\":\"header\",\"data\":{\"text\":\"Environmental feedback is what makes the loop useful\",\"level\":2},\"tunes\":{}},{\"id\":\"p-feedback-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"An agent becomes operationally meaningful when it can observe whether its action worked. Tool output, test results, API responses, filesystem state, browser state and application records provide external evidence that the system can use to revise its next decision.\"},\"tunes\":{}},{\"id\":\"p-feedback-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Anthropic's agent guidance emphasizes this feedback loop: agents use tools, obtain ground truth from the environment, assess progress and continue or request human input.\"},\"tunes\":{}},{\"id\":\"feedback-rule\",\"type\":\"callout\",\"data\":{\"variant\":\"warning\",\"title\":\"Self-report is not environmental proof\",\"body\":\"An agent saying “the task is complete” does not prove completion. Where possible, verify the final state through an external system, test, file, transaction record or other observable outcome.\"},\"tunes\":{}},{\"id\":\"h-state\",\"type\":\"header\",\"data\":{\"text\":\"Agent state is not the same as model context\",\"level\":2},\"tunes\":{}},{\"id\":\"p-state-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A long-running task may need state that cannot or should not remain in the model context: task IDs, checkpoints, artifacts, approvals, external object identifiers, retry counters and workflow status.\"},\"tunes\":{}},{\"id\":\"p-state-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The runtime can preserve this durable state outside the model window and reconstruct the context required for the next step. This keeps model-visible context focused while maintaining continuity and resumability.\"},\"tunes\":{}},{\"id\":\"h-memory\",\"type\":\"header\",\"data\":{\"text\":\"Memory is optional, not the definition of an agent\",\"level\":2},\"tunes\":{}},{\"id\":\"p-memory-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"An agent can operate successfully without long-term memory if the complete task fits inside one bounded run. Memory becomes useful when information must persist across sessions, tasks or long execution horizons.\"},\"tunes\":{}},{\"id\":\"p-memory-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"RAG, memory, state and context solve different problems. Treating a vector database as “the agent memory” or conversation history as “the state machine” usually hides important lifecycle and authority boundaries.\"},\"tunes\":{}},{\"id\":\"ref-memory\",\"type\":\"referralArticle\",\"data\":{\"url\":\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context\",\"title\":\"AI Agent Memory Is Not RAG: How to Separate Memory, Retrieval, State and Context\",\"excerpt\":\"A practical architecture separating persistent memory, authoritative application state, retrieval and the context supplied to the model.\",\"ctaLabel\":\"Read the memory architecture article\"},\"tunes\":{}},{\"id\":\"h-context\",\"type\":\"header\",\"data\":{\"text\":\"Context engineering becomes dynamic in agents\",\"level\":2},\"tunes\":{}},{\"id\":\"p-context-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Every tool call can produce new context. Every step can also make earlier information obsolete. A strong agent runtime therefore rebuilds or curates context as execution progresses rather than replaying everything indefinitely.\"},\"tunes\":{}},{\"id\":\"p-context-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Tool definitions, task state, retrieved evidence, observations and memory all compete for the model's attention. Long-running agents need trimming, compaction or just-in-time loading so context remains relevant to the current decision.\"},\"tunes\":{}},{\"id\":\"h-sideeffects\",\"type\":\"header\",\"data\":{\"text\":\"Read tools and side-effect tools have different risk\",\"level\":2},\"tunes\":{}},{\"id\":\"sideeffect-comparison\",\"type\":\"comparison\",\"data\":{\"title\":\"Information access versus external action\",\"layout\":\"table\",\"columns\":[{\"id\":\"read\",\"label\":\"Read \u002F observe\"},{\"id\":\"write\",\"label\":\"Write \u002F act\"}],\"rows\":[{\"id\":\"examples\",\"label\":\"Examples\",\"values\":[\"\",\"\"]},{\"id\":\"mainrisk\",\"label\":\"Main risk\",\"values\":[\"\",\"\"]},{\"id\":\"control\",\"label\":\"Typical control\",\"values\":[\"\",\"\"]}]},\"tunes\":{}},{\"id\":\"h-approval\",\"type\":\"header\",\"data\":{\"text\":\"Human-in-the-loop is a control mechanism, not the opposite of agentic AI\",\"level\":2},\"tunes\":{}},{\"id\":\"p-approval-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"An agent does not stop being agentic because a human approves consequential steps. The model can still autonomously inspect, reason, search and prepare an action while the runtime requires human confirmation before execution.\"},\"tunes\":{}},{\"id\":\"p-approval-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"OpenAI's current agent safety guidance explicitly recommends approvals for tool operations in higher-risk workflows. Anthropic likewise emphasizes checkpoints and human judgment where agents encounter blockers or consequential decisions.\"},\"tunes\":{}},{\"id\":\"p-approval-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The useful architecture question is not “human or autonomous?” but which decisions can be delegated, which require review and which must remain deterministic?\"},\"tunes\":{}},{\"id\":\"h-stopping\",\"type\":\"header\",\"data\":{\"text\":\"Agents need explicit stopping conditions\",\"level\":2},\"tunes\":{}},{\"id\":\"stop-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Stopping condition\",\"Purpose\"],[\"Successful verified outcome\",\"End when the external target state is confirmed.\"],[\"Maximum steps\",\"Prevent runaway loops.\"],[\"Time budget\",\"Bound wall-clock execution.\"],[\"Cost\u002Ftoken budget\",\"Limit resource consumption.\"],[\"Repeated-action detector\",\"Stop loops that are no longer making progress.\"],[\"Permission boundary\",\"Pause or stop when the next required action is not permitted.\"],[\"Human approval checkpoint\",\"Wait before consequential execution.\"],[\"Unrecoverable tool failure\",\"Escalate instead of retrying indefinitely.\"],[\"Uncertainty threshold\",\"Ask for clarification when the task cannot be safely inferred.\"]]},\"tunes\":{}},{\"id\":\"h-errors\",\"type\":\"header\",\"data\":{\"text\":\"Recovery is part of agent behavior\",\"level\":2},\"tunes\":{}},{\"id\":\"p-errors-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Agents operate in environments that fail: APIs time out, files change, credentials expire, webpages move and tools return malformed output. A useful agentic system therefore needs recovery behavior, not just a happy-path tool loop.\"},\"tunes\":{}},{\"id\":\"p-errors-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Recovery can include retry with limits, choosing another tool, re-reading current state, asking the user, rolling back a partial action or escalating to a human.\"},\"tunes\":{}},{\"id\":\"p-errors-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Retries also need idempotency awareness. Repeating a read is usually low risk; repeating a payment or message send can create duplicate side effects.\"},\"tunes\":{}},{\"id\":\"h-single-multi\",\"type\":\"header\",\"data\":{\"text\":\"Agentic AI does not require multiple agents\",\"level\":2},\"tunes\":{}},{\"id\":\"p-multi-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A single agent with a clear tool set is often simpler and easier to evaluate than a multi-agent architecture. Multiple agents are useful when specialization materially improves tool isolation, policy isolation, prompt clarity, ownership or trace legibility.\"},\"tunes\":{}},{\"id\":\"p-multi-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"OpenAI's current orchestration guidance explicitly recommends starting with one agent where possible and adding specialists only when the contract or ownership boundary materially changes.\"},\"tunes\":{}},{\"id\":\"p-multi-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Multi-agent systems add new problems: delegation quality, duplicated context, conflicting state, handoff semantics, identity, cost and distributed failure handling.\"},\"tunes\":{}},{\"id\":\"h-protocols\",\"type\":\"header\",\"data\":{\"text\":\"Agent protocols are interoperability layers, not the agent itself\",\"level\":2},\"tunes\":{}},{\"id\":\"p-protocols-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Protocols such as MCP and A2A can make an agent architecture interoperable, but they do not create the agent loop by themselves. MCP can expose tools and resources. A2A can connect independently implemented agents. The application still needs runtime, authorization, state, evaluation and domain logic.\"},\"tunes\":{}},{\"id\":\"p-protocols-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"This is why protocol capability must remain separate from business authority. Discovering a tool through MCP does not prove the current principal is allowed to use it. Receiving a task through A2A does not prove the remote agent may perform every requested action.\"},\"tunes\":{}},{\"id\":\"ref-protocols\",\"type\":\"referralArticle\",\"data\":{\"url\":\"https:\u002F\u002Fstajic.de\u002Fde\u002Fblog\u002Fmcp-vs-a2a-vs-ucp-vs-ap2-vs-a2ui-the-agent-protocol-stack-explained\",\"title\":\"MCP vs A2A vs UCP vs AP2 vs A2UI: The Agent Protocol Stack Explained\",\"excerpt\":\"A protocol-responsibility map showing why tool access, agent collaboration, commerce, payment authority and agent-driven UI belong to different interoperability boundaries.\",\"ctaLabel\":\"Read the agent protocol stack\"},\"tunes\":{}},{\"id\":\"h-reliability\",\"type\":\"header\",\"data\":{\"text\":\"The trajectory is part of agent reliability\",\"level\":2},\"tunes\":{}},{\"id\":\"p-rel-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A final answer is insufficient evidence for an agentic system because an agent can reach the right result through an unsafe or invalid path. It may use an unauthorized tool, skip a required check, retry a side effect, rely on stale state or accidentally succeed.\"},\"tunes\":{}},{\"id\":\"p-rel-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Evaluation therefore needs execution traces: decisions, tool calls, approvals, observations, state changes and final outcome. Current OpenAI safety guidance recommends trace graders and evals; Anthropic's 2026 agent-evaluation guidance similarly treats multi-turn tool trajectories as first-class evaluation objects.\"},\"tunes\":{}},{\"id\":\"p-rel-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The stronger reliability question is: Did the agent reach an acceptable outcome through an acceptable, recoverable and auditable trajectory?\"},\"tunes\":{}},{\"id\":\"ref-reliability\",\"type\":\"referralArticle\",\"data\":{\"url\":\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fai-agent-reliability-why-the-final-answer-is-not-enough\",\"title\":\"AI Agent Reliability: Why the Final Answer Is Not Enough\",\"excerpt\":\"Why production evaluation must inspect trajectories, tool use, state transitions and recoverability rather than only final answers.\",\"ctaLabel\":\"Read the reliability article\"},\"tunes\":{}},{\"id\":\"h-security\",\"type\":\"header\",\"data\":{\"text\":\"Agentic systems increase the security surface\",\"level\":2},\"tunes\":{}},{\"id\":\"security-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Risk\",\"Why agents amplify it\",\"Architecture response\"],[\"Prompt injection\",\"Untrusted content can influence future tool decisions\",\"Separate instructions from data; constrain tools; sanitize or structure external input where possible\"],[\"Excessive permissions\",\"Reasoning errors can become real side effects\",\"Least privilege, scoped credentials, per-tool policy and approvals\"],[\"Credential exposure\",\"Tools may need powerful secrets\",\"Keep secrets outside model context; broker access through trusted runtime\"],[\"Confused deputy\",\"Agent may act with authority broader than the requesting user\",\"Bind execution to user\u002Fservice identity and re-authorize consequential actions\"],[\"Runaway loops\",\"Model repeatedly calls tools without progress\",\"Step, time and cost budgets plus loop detection\"],[\"State drift\",\"Environment changes after the agent formed a plan\",\"Re-read authoritative state before consequential actions\"],[\"Indirect injection\",\"Tool\u002Fweb\u002Fdocument content contains instructions aimed at the model\",\"Treat external content as untrusted data, not instruction authority\"],[\"Audit gap\",\"Final result cannot show what was executed\",\"Trace tool calls, approvals, identities and state changes\"]]},\"tunes\":{}},{\"id\":\"h-observability\",\"type\":\"header\",\"data\":{\"text\":\"Agent observability must follow the loop\",\"level\":2},\"tunes\":{}},{\"id\":\"p-obs-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Traditional service observability records requests, latency and errors. Agent observability needs an additional execution model: which agent was active, which model version made the decision, what context was available, which tool was selected, what arguments were sent, what result came back and why execution stopped.\"},\"tunes\":{}},{\"id\":\"p-obs-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"For sensitive systems, traces themselves require access control and retention policy because prompts, tool outputs and artifacts can contain confidential data.\"},\"tunes\":{}},{\"id\":\"h-eval\",\"type\":\"header\",\"data\":{\"text\":\"How to evaluate an agentic system\",\"level\":2},\"tunes\":{}},{\"id\":\"eval-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Dimension\",\"Question\",\"Example evidence\"],[\"Task success\",\"Did the requested outcome occur?\",\"External state, tests, business outcome\"],[\"Trajectory quality\",\"Were the steps acceptable?\",\"Tool\u002Faction trace\"],[\"Tool selection\",\"Did the agent choose appropriate capabilities?\",\"Expected vs actual tool calls\"],[\"Permission adherence\",\"Did it stay inside allowed authority?\",\"Authorization logs and denied-action tests\"],[\"State handling\",\"Did it use current authoritative state?\",\"Freshness checks and state-change tests\"],[\"Recovery\",\"Did it respond correctly to failures?\",\"Injected timeout\u002Ferror scenarios\"],[\"Stopping behavior\",\"Did it stop at the right point?\",\"Step counts, loop detection, final-state proof\"],[\"Human escalation\",\"Did it ask when review was required?\",\"Approval\u002Fescalation traces\"],[\"Cost\u002Flatency\",\"Was autonomy worth the operational cost?\",\"Tokens, tool calls, duration\"],[\"Robustness\",\"Does it survive realistic environment variation?\",\"Repeated and adversarial trials\"]]},\"tunes\":{}},{\"id\":\"h-use\",\"type\":\"header\",\"data\":{\"text\":\"When an agent is appropriate\",\"level\":2},\"tunes\":{}},{\"id\":\"use-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Use an agent when\",\"Prefer a workflow or simple call when\"],[\"The number or order of steps cannot be known reliably in advance\",\"The sequence is stable and deterministic\"],[\"The system must inspect the environment and adapt\",\"A single retrieval + generation step is sufficient\"],[\"Several tools may be useful depending on intermediate results\",\"One known API call solves the task\"],[\"The task benefits from iterative verification or repair\",\"The answer can be produced directly from supplied context\"],[\"Failures require flexible recovery behavior\",\"Failure branches are simple and can be encoded explicitly\"],[\"Human review can be inserted at meaningful checkpoints\",\"Every step is high-risk and must be manually controlled anyway\"],[\"Expected value justifies extra latency, cost and complexity\",\"Predictability and low cost matter more than flexibility\"]]},\"tunes\":{}},{\"id\":\"p-use-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"A strong default is to start with the simplest solution that works and increase agentic complexity only when flexibility produces measurable value. Agents trade predictability, latency and cost for adaptive execution.\"},\"tunes\":{}},{\"id\":\"h-implementation\",\"type\":\"header\",\"data\":{\"text\":\"Original implementation evidence\",\"level\":2},\"tunes\":{}},{\"id\":\"h-client\",\"type\":\"header\",\"data\":{\"text\":\"Aaasaasa AI Client: model, runtime and permission are separate\",\"level\":3},\"tunes\":{}},{\"id\":\"p-client-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Aaasaasa AI Client explicitly separates agent\u002Fclient, provider, model, runtime location and permissions. Its architecture documentation treats permissions as central tool\u002Fworkspace policy rather than a model property.\"},\"tunes\":{}},{\"id\":\"p-client-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The same application can expose Direct Chat with no filesystem or shell tools while a Codex runtime operates under a selected workspace and permission profile. This demonstrates a core agentic architecture boundary: changing the runtime\u002Ftool surface changes what the system can do even when model access remains available.\"},\"tunes\":{}},{\"id\":\"p-client-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The repository also distinguishes a local Codex runtime from model location: a local runtime can call a cloud model. This prevents the common mistake of equating “agent runs locally” with “inference is local.”\"},\"tunes\":{}},{\"id\":\"p-client-4\",\"type\":\"paragraph\",\"data\":{\"text\":\"The implementation disables embedded execution paths whose approval semantics do not satisfy the required permission model. This supports the principle that agent capability should not bypass runtime authorization simply because an underlying framework can execute tools.\"},\"tunes\":{}},{\"id\":\"h-sot-agent\",\"type\":\"header\",\"data\":{\"text\":\"Source of Truth Research Engine: bounded agentic research stages\",\"level\":3},\"tunes\":{}},{\"id\":\"p-sot-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"The Source of Truth Research Engine uses a bounded research pipeline: discover → acquire → extract → verify → contradict → synthesize. Research jobs can execute through an AI runtime while evidence, sources, claims and contradictions remain in an external persistent store.\"},\"tunes\":{}},{\"id\":\"p-sot-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"This is intentionally more controlled than an unconstrained autonomous research agent. The stages provide guardrails around what kind of work should happen next while still allowing model-driven research inside each bounded task.\"},\"tunes\":{}},{\"id\":\"p-sot-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"That distinction is useful evidence for agent design: autonomy can be placed inside a structured delivery envelope rather than applied uniformly to the entire process.\"},\"tunes\":{}},{\"id\":\"impl-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Implemented pattern\",\"Agentic architecture lesson\"],[\"Direct Chat has no OS tools\",\"A model can exist without agentic execution capability.\"],[\"Codex runtime has workspace permission profile\",\"Tool authority belongs to runtime policy, not model capability.\"],[\"Provider\u002Fmodel\u002Fruntime are separate concepts\",\"Agent harness location and inference location are independent decisions.\"],[\"Permission broker for tool-capable runtimes\",\"Capability exposure can be centralized and governed.\"],[\"Bounded research stages\",\"Autonomy can operate inside explicit process boundaries.\"],[\"Persistent claims\u002Fevidence outside model context\",\"Agent state and evidence do not need to live only in conversation history.\"]]},\"tunes\":{}},{\"id\":\"impl-boundary\",\"type\":\"callout\",\"data\":{\"variant\":\"note\",\"title\":\"Evidence boundary\",\"body\":\"These projects demonstrate concrete agent\u002Fruntime, permission and bounded-research patterns. They are not presented as proof of large-scale commercial autonomous-agent deployment.\"},\"tunes\":{}},{\"id\":\"h-failures\",\"type\":\"header\",\"data\":{\"text\":\"Common agentic AI failure modes\",\"level\":2},\"tunes\":{}},{\"id\":\"failure-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Failure mode\",\"What actually failed\"],[\"“Agent” is only a chatbot with tools listed in the prompt\",\"No reliable runtime loop or tool execution architecture exists\"],[\"Tool support is treated as permission\",\"Capability and authorization boundaries are collapsed\"],[\"Agent trusts its own completion statement\",\"Outcome is not verified against external state\"],[\"Every task becomes multi-agent\",\"Complexity increases without a real ownership or specialization boundary\"],[\"Conversation history is used as durable state\",\"Resumability and authoritative state become fragile\"],[\"Agent retries side effects blindly\",\"Duplicate messages, payments or state changes become possible\"],[\"No step\u002Fcost limits\",\"Agent can loop indefinitely or consume uncontrolled resources\"],[\"Tool output is trusted as instruction\",\"Indirect prompt injection can redirect behavior\"],[\"Correct final answer is the only evaluation\",\"Unsafe or invalid trajectories remain invisible\"],[\"Model upgrade is treated as transparent\",\"Tool selection, planning and stopping behavior can change\"],[\"One broad tool exposes many privileged operations\",\"Blast radius increases and intent becomes harder to validate\"],[\"Human approval exists but reviewer lacks context\",\"Approval becomes ceremonial rather than effective\"]]},\"tunes\":{}},{\"id\":\"h-misconceptions\",\"type\":\"header\",\"data\":{\"text\":\"Common misconceptions\",\"level\":2},\"tunes\":{}},{\"id\":\"misconceptions-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Misconception\",\"Correction\"],[\"“An LLM is an agent.”\",\"The model is the decision component; the agent is the surrounding system that manages tools, state and iteration.\"],[\"“Tool calling automatically means agentic AI.”\",\"A single bounded tool call may not involve an adaptive multi-step agent loop.\"],[\"“Agents must be fully autonomous.”\",\"Agentic systems can require approvals and operate under narrow permission boundaries.\"],[\"“Agents need long-term memory.”\",\"Memory is optional; many useful agents complete bounded tasks without cross-session memory.\"],[\"“Agents must create a written plan first.”\",\"Planning can be explicit or implicit and can occur one step at a time.\"],[\"“Multi-agent is more advanced than single-agent.”\",\"It is more complex; use it only when specialization or ownership boundaries justify it.\"],[\"“MCP creates an agent.”\",\"MCP exposes tools\u002Fresources; the runtime still needs an agent loop and authorization model.\"],[\"“A local runtime means the model is local.”\",\"Runtime location and inference\u002Fprovider location are separate.\"],[\"“If the final result is correct, the agent worked correctly.”\",\"An unsafe or unauthorized trajectory can still produce a correct result.\"],[\"“Human approval removes autonomy.”\",\"Approval can constrain selected actions while the rest of the process remains model-directed.\"]]},\"tunes\":{}},{\"id\":\"h-design\",\"type\":\"header\",\"data\":{\"text\":\"A practical agent design sequence\",\"level\":2},\"tunes\":{}},{\"id\":\"design-flow\",\"type\":\"processFlow\",\"data\":{\"title\":\"Design the agent from authority outward\",\"orientation\":\"auto\",\"steps\":[{\"label\":\"1. Define the outcome\",\"description\":\"State what external result or artifact proves task success.\"},{\"label\":\"2. Decide whether an agent is actually needed\",\"description\":\"Prefer a simple call or deterministic workflow when the path is predictable.\"},{\"label\":\"3. Identify state and Source of Truth\",\"description\":\"Define which systems own current facts, task progress and business state.\"},{\"label\":\"4. Define the tool surface\",\"description\":\"Expose the smallest set of clear capabilities required for the task.\"},{\"label\":\"5. Bind identity and permissions\",\"description\":\"Separate user authority, agent\u002Fruntime permissions and tool capabilities.\"},{\"label\":\"6. Choose autonomy boundaries\",\"description\":\"Specify what the model may decide dynamically and what remains deterministic.\"},{\"label\":\"7. Add approval checkpoints\",\"description\":\"Require review before consequential or irreversible actions where appropriate.\"},{\"label\":\"8. Define stopping and recovery\",\"description\":\"Set success proof, budgets, timeouts, retries, escalation and loop controls.\"},{\"label\":\"9. Design context\u002Fstate management\",\"description\":\"Keep current state, memory, tool observations and durable artifacts in the correct layers.\"},{\"label\":\"10. Trace the trajectory\",\"description\":\"Record enough execution structure to debug and audit model\u002Ftool decisions.\"},{\"label\":\"11. Evaluate realistic failures\",\"description\":\"Test stale state, tool errors, prompt injection, ambiguous requests and changed environments.\"},{\"label\":\"12. Expand autonomy only from evidence\",\"description\":\"Increase permissions or execution horizon when evaluation shows the benefit justifies the risk.\"}]},\"tunes\":{}},{\"id\":\"h-checklist\",\"type\":\"header\",\"data\":{\"text\":\"Agentic AI architecture checklist\",\"level\":2},\"tunes\":{}},{\"id\":\"checklist-table\",\"type\":\"table\",\"data\":{\"withHeadings\":true,\"stretched\":false,\"content\":[[\"Question\",\"Expected evidence\"],[\"What proves success?\",\"External outcome, artifact, test or authoritative state.\"],[\"Why is an agent needed?\",\"The path genuinely depends on intermediate observations.\"],[\"Which decisions are model-driven?\",\"Explicit autonomy boundary.\"],[\"Which tools exist?\",\"Small, documented, unambiguous capability set.\"],[\"Who may use each tool?\",\"Identity- and context-aware authorization policy.\"],[\"Which actions need approval?\",\"Consequence-based review rules.\"],[\"Where does task state live?\",\"Application-owned state separate from transient model context.\"],[\"How does the agent recover?\",\"Retry, re-read, rollback, clarification and escalation behavior.\"],[\"How does it stop?\",\"Verified completion plus step\u002Ftime\u002Fcost limits.\"],[\"How are side effects protected?\",\"Validation, idempotency, least privilege and confirmation.\"],[\"Can execution be reconstructed?\",\"Tool, approval and state-transition traces.\"],[\"How is it evaluated?\",\"Outcome + trajectory + robustness tests.\"],[\"What changes after a model\u002Fruntime update?\",\"Regression suite for tool selection, permissions, stopping and recovery.\"]]},\"tunes\":{}},{\"id\":\"h-edge\",\"type\":\"header\",\"data\":{\"text\":\"Edge cases and limitations\",\"level\":2},\"tunes\":{}},{\"id\":\"p-edge-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Some systems are “agentic” only in a narrow routing sense: the model selects one specialist or tool and then the rest of the workflow is deterministic. That can still be useful, but it should not be described as equivalent to a long-running autonomous agent.\"},\"tunes\":{}},{\"id\":\"p-edge-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Highly consequential domains may intentionally restrict agent autonomy. An AI system can inspect evidence, prepare recommendations and fill structured forms while a human remains the only actor allowed to commit the final transaction.\"},\"tunes\":{}},{\"id\":\"p-edge-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Some environments are well suited to agents because feedback is objective. Coding agents can run tests; infrastructure agents can inspect metrics; data agents can validate query results. Open-ended domains with weak feedback require more cautious evaluation.\"},\"tunes\":{}},{\"id\":\"p-edge-4\",\"type\":\"paragraph\",\"data\":{\"text\":\"An agent can operate entirely locally, entirely through managed cloud services or in a hybrid architecture. Agentic behavior describes control flow, not hosting location.\"},\"tunes\":{}},{\"id\":\"p-edge-5\",\"type\":\"paragraph\",\"data\":{\"text\":\"The term “reasoning” should not be used as proof that the agent's internal process is correct. Production assurance should rely on observable inputs, actions, outputs, state and evaluation rather than unverifiable claims about hidden reasoning.\"},\"tunes\":{}},{\"id\":\"h-change\",\"type\":\"header\",\"data\":{\"text\":\"What would change this answer?\",\"level\":2},\"tunes\":{}},{\"id\":\"p-change-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Vendor APIs and agent frameworks will continue to evolve, but the architecture boundary is stable: a model proposes decisions, a runtime manages the loop, tools connect to the environment, permissions constrain actions and external observations determine what actually happened.\"},\"tunes\":{}},{\"id\":\"p-change-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"As models become more reliable, systems may safely delegate longer horizons or more complex recovery behavior. As runtime verification and authorization improve, some approval steps may become automated. Those are changes in autonomy level, not changes to the fundamental responsibility layers.\"},\"tunes\":{}},{\"id\":\"p-change-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The recommended architecture also changes by consequence. A research agent that only reads public sources can tolerate different controls from an agent that writes production configuration or moves money.\"},\"tunes\":{}},{\"id\":\"h-related\",\"type\":\"header\",\"data\":{\"text\":\"Related canonical knowledge\",\"level\":2},\"tunes\":{}},{\"id\":\"p-related-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Agentic AI sits above several prerequisite layers: context engineering determines what the model sees; Source-of-Truth architecture determines which information is authoritative; retrieval supplies external evidence; runtime architecture determines what can execute.\"},\"tunes\":{}},{\"id\":\"p-related-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"Downstream nodes include tool calling, MCP, A2A, agent identity, permissions, auditability, human-in-the-loop, orchestration, memory and multi-agent systems.\"},\"tunes\":{}},{\"id\":\"p-related-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The protocol stack article should therefore be read after the basic agent concept: protocols standardize boundaries around agents; they do not define agentic behavior itself.\"},\"tunes\":{}},{\"id\":\"h-faq\",\"type\":\"header\",\"data\":{\"text\":\"Frequently asked questions\",\"level\":2},\"tunes\":{}},{\"id\":\"faq\",\"type\":\"faq\",\"data\":{\"title\":\"Agentic AI FAQ\",\"items\":[{\"id\":\"faq1\",\"question\":\"What is agentic AI?\",\"answer\":\"Agentic AI is an AI system in which a model can pursue a goal over multiple steps by choosing actions or tools, observing results, updating its state and continuing until a stopping condition is reached.\"},{\"id\":\"faq2\",\"question\":\"What is the difference between an LLM and an AI agent?\",\"answer\":\"An LLM produces outputs from inputs. An agent combines a model with a runtime, tools, state, permissions, context management and an iterative execution loop.\"},{\"id\":\"faq3\",\"question\":\"Does tool calling make a system an agent?\",\"answer\":\"Not necessarily. A single tool-assisted model response can be bounded and non-agentic. Agentic behavior appears when tool observations drive an adaptive multi-step loop.\"},{\"id\":\"faq4\",\"question\":\"What is the difference between an agent and an AI workflow?\",\"answer\":\"A workflow usually follows a process path defined in application code. An agent has more model-driven control over which steps and tools to use based on intermediate observations.\"},{\"id\":\"faq5\",\"question\":\"Do agents need memory?\",\"answer\":\"No. Long-term memory is useful for persistent information across sessions, but many agents complete bounded tasks using only current task state and context.\"},{\"id\":\"faq6\",\"question\":\"Do AI agents need multiple agents?\",\"answer\":\"No. A single agent is often simpler. Multi-agent systems are justified when specialization, tool isolation, policy isolation or ownership boundaries materially improve the system.\"},{\"id\":\"faq7\",\"question\":\"Can an agent be human-in-the-loop?\",\"answer\":\"Yes. The agent can autonomously perform low-risk analysis and preparation while the runtime pauses for human approval before consequential actions.\"},{\"id\":\"faq8\",\"question\":\"Is MCP an agent framework?\",\"answer\":\"No. MCP is an interoperability protocol for exposing tools, resources and prompts. An agent runtime can use MCP, but still needs its own loop, state, authorization and evaluation.\"},{\"id\":\"faq9\",\"question\":\"How do you know an agent actually completed a task?\",\"answer\":\"Where possible, verify success through external state, tests, artifacts or authoritative system records rather than trusting the model's own completion statement.\"}]},\"tunes\":{}},{\"id\":\"h-glossary\",\"type\":\"header\",\"data\":{\"text\":\"Glossary\",\"level\":2},\"tunes\":{}},{\"id\":\"glossary\",\"type\":\"glossary\",\"data\":{\"title\":\"Key agentic AI terms\",\"entries\":[{\"term\":\"Agentic AI\",\"definition\":\"AI system behavior in which a model dynamically directs multi-step execution using tools, observations and state toward a goal.\",\"anchor\":\"agentic-ai\"},{\"term\":\"AI agent\",\"definition\":\"A model-centered system with runtime, tools, state and an execution loop that can pursue a task over multiple steps.\",\"anchor\":\"ai-agent\"},{\"term\":\"Agent loop\",\"definition\":\"Repeated cycle of model decision, tool\u002Faction execution, observation and updated model decision until stopping.\",\"anchor\":\"agent-loop\"},{\"term\":\"Runtime \u002F harness\",\"definition\":\"The execution layer that manages the model loop, tools, state, approvals, context, errors and stopping conditions.\",\"anchor\":\"runtime-harness\"},{\"term\":\"Tool\",\"definition\":\"A capability exposed to the model for reading information, computing, delegating or changing external state.\",\"anchor\":\"tool\"},{\"term\":\"Observation\",\"definition\":\"Information returned from a tool or environment and supplied to a later agent step.\",\"anchor\":\"observation\"},{\"term\":\"Agent state\",\"definition\":\"Persistent task or execution information that exists outside a single model output and may survive across steps or pauses.\",\"anchor\":\"agent-state\"},{\"term\":\"Autonomy boundary\",\"definition\":\"The explicit limit defining which decisions and actions the model may control dynamically.\",\"anchor\":\"autonomy-boundary\"},{\"term\":\"Human-in-the-loop\",\"definition\":\"A control pattern in which human review, input or approval is required at selected points in an AI-driven process.\",\"anchor\":\"human-in-the-loop\"},{\"term\":\"Trajectory\",\"definition\":\"The sequence of relevant states, decisions, tool calls, actions and observations between task request and final outcome.\",\"anchor\":\"trajectory\"},{\"term\":\"Idempotency\",\"definition\":\"Property that allows an operation to be repeated without unintentionally applying the same side effect multiple times.\",\"anchor\":\"idempotency\"}]},\"tunes\":{}},{\"id\":\"h-conclusion\",\"type\":\"header\",\"data\":{\"text\":\"Conclusion\",\"level\":2},\"tunes\":{}},{\"id\":\"p-conclusion-1\",\"type\":\"paragraph\",\"data\":{\"text\":\"Agentic AI is not simply a smarter model or a chatbot with more tools. It is a system architecture in which a model participates in an iterative control loop: decide, act, observe, update and continue.\"},\"tunes\":{}},{\"id\":\"p-conclusion-2\",\"type\":\"paragraph\",\"data\":{\"text\":\"The model provides flexible decision making, but the surrounding runtime must own execution reality: permissions, tool access, state, approvals, retries, budgets, stopping conditions, tracing and verification.\"},\"tunes\":{}},{\"id\":\"p-conclusion-3\",\"type\":\"paragraph\",\"data\":{\"text\":\"The most useful design principle is therefore: delegate tactical choice to the model only inside explicit technical and business boundaries. Agentic capability becomes production capability only when autonomy, authority and evidence remain separable.\"},\"tunes\":{}},{\"id\":\"h-sources\",\"type\":\"header\",\"data\":{\"text\":\"Primary sources and current guidance\",\"level\":2},\"tunes\":{}},{\"id\":\"p-sources-note\",\"type\":\"paragraph\",\"data\":{\"text\":\"The sources below support the current architectural distinctions around agents, workflows, loops, tools, orchestration, safety and evaluation. Project sections are original implementation evidence and are explicitly bounded to what the repositories demonstrate.\"},\"tunes\":{}},{\"id\":\"src-openai-agents\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"OpenAI — Agents\",\"description\":\"Current developer guidance defining runtime choices for multi-step work, tools, state, orchestration and agent execution.\"}},\"tunes\":{}},{\"id\":\"src-openai-definitions\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\u002Fdefine-agents\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"OpenAI — Agent definitions\",\"description\":\"Current documentation describing an agent as a model plus instructions and optional runtime behavior including tools, guardrails, MCP servers and handoffs.\"}},\"tunes\":{}},{\"id\":\"src-openai-running\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\u002Frunning-agents\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"OpenAI — Running agents\",\"description\":\"Current documentation of the agent loop: model call, tool execution or handoff, continuation and final stopping point.\"}},\"tunes\":{}},{\"id\":\"src-openai-orchestration\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagents\u002Forchestration\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"OpenAI — Orchestration and handoffs\",\"description\":\"Current guidance on handoffs, agents-as-tools and when specialist agents add useful ownership or capability boundaries.\"}},\"tunes\":{}},{\"id\":\"src-openai-safety\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fdevelopers.openai.com\u002Fapi\u002Fdocs\u002Fguides\u002Fagent-builder-safety\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"OpenAI — Safety in building agents\",\"description\":\"Current safety guidance covering tool approvals, prompt injection, guardrails and trace-based evaluation.\"}},\"tunes\":{}},{\"id\":\"src-anthropic-agents\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fbuilding-effective-agents\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Anthropic — Building effective agents\",\"description\":\"Engineering guidance distinguishing predefined workflows from model-directed agents and describing tool-based environmental feedback loops.\"}},\"tunes\":{}},{\"id\":\"src-anthropic-context\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Feffective-context-engineering-for-ai-agents\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Anthropic — Effective context engineering for AI agents\",\"description\":\"Practical framing of agents as LLMs autonomously using tools in a loop, with dynamic just-in-time context management.\"}},\"tunes\":{}},{\"id\":\"src-anthropic-evals\",\"type\":\"linkTool\",\"data\":{\"link\":\"https:\u002F\u002Fwww.anthropic.com\u002Fengineering\u002Fdemystifying-evals-for-ai-agents\",\"meta\":{\"image\":{\"url\":\"\"},\"title\":\"Anthropic — Demystifying evals for AI agents\",\"description\":\"2026 guidance on evaluating multi-turn agents that call tools, modify state and adapt to intermediate results.\"}},\"tunes\":{}}],\"version\":\"2.31.6\"}",{"time":1579,"blocks":1580,"version":2701},1791487185746,[1581,1585,1590,1595,1600,1604,1608,1612,1616,1620,1624,1628,1632,1636,1662,1666,1670,1674,1678,1699,1703,1707,1739,1743,1747,1789,1793,1797,1801,1806,1810,1814,1818,1822,1826,1854,1858,1862,1866,1870,1874,1878,1882,1886,1890,1894,1898,1903,1907,1911,1915,1919,1923,1927,1934,1938,1942,1946,1950,1969,1973,1977,1981,1985,1989,2023,2027,2031,2035,2039,2043,2047,2051,2055,2059,2063,2067,2073,2077,2081,2085,2089,2096,2100,2139,2143,2147,2151,2155,2202,2206,2234,2238,2242,2246,2250,2254,2258,2262,2266,2270,2274,2278,2303,2308,2312,2355,2359,2396,2400,2441,2445,2490,2494,2498,2502,2506,2510,2514,2518,2522,2526,2530,2534,2538,2542,2546,2550,2582,2586,2621,2625,2629,2633,2637,2641,2645,2652,2659,2666,2673,2680,2687,2694],{"id":215,"data":1582,"type":218,"tunes":1584},{"text":1583},"Agentic AI is an AI system in which a model can pursue a goal across multiple steps by deciding what to do next, using tools or other capabilities, observing the results, updating its working state and continuing until it reaches a stopping condition. The model alone is not the agent. A usable agent also needs a runtime or harness that manages context, tool execution, state, permissions, approvals, errors and the loop between decisions and observations.",{},{"id":221,"data":1586,"type":226,"tunes":1589},{"body":1587,"title":1588,"variant":225},"A normal model call is usually \u003Cstrong>input → model → output\u003C\u002Fstrong>. An agentic system is closer to \u003Cstrong>goal → decision → tool\u002Faction → observation → updated decision → … → result\u003C\u002Fstrong>.\u003Cbr>\u003Cbr>The critical distinction is not whether an application uses an LLM or function calling. It is whether the system gives the model meaningful control over the next step of a multi-step process while a runtime constrains what the model is actually allowed to do.","Direct answer",{},{"id":229,"data":1591,"type":226,"tunes":1594},{"body":1592,"title":1593,"variant":233},"A model may know how to call a tool. A runtime may expose that tool. Neither fact means the current user or agent is authorized to execute the underlying business action. \u003Cstrong>Tool capability, tool permission and business authority are separate layers.\u003C\u002Fstrong>","Capability is not authority",{},{"id":236,"data":1596,"type":226,"tunes":1599},{"body":1597,"title":1598,"variant":240},"Agent terminology still varies across vendors and research communities. OpenAI currently defines agent runtimes around multi-step work, tools, state and orchestration. Anthropic's practical distinction remains useful: workflows follow predefined code paths, while agents dynamically direct their own process and tool use. This article therefore treats “agentic AI” as an architectural spectrum rather than one standardized product category.","Current-source note — 8 October 2026",{},{"id":243,"data":1601,"type":248,"tunes":1603},{"title":1602,"maxLevel":246,"minLevel":247},"Contents",{},{"id":251,"data":1605,"type":42,"tunes":1607},{"text":1606,"level":247},"What agentic AI really means",{},{"id":256,"data":1609,"type":218,"tunes":1611},{"text":1610},"The important shift from ordinary generative AI to agentic AI is control over process. A normal assistant can answer a question using the context it receives. An agent can decide that answering requires additional steps: inspect a file, search a repository, query an API, ask for clarification, run a test, update a ticket, delegate a subtask or retry after a failed action.",{},{"id":261,"data":1613,"type":218,"tunes":1615},{"text":1614},"This does not require unlimited autonomy. An agent can operate inside a narrow sandbox, under strict permissions, with approval required before every consequential action. The system is still agentic if the model dynamically chooses among permitted next steps.",{},{"id":266,"data":1617,"type":218,"tunes":1619},{"text":1618},"The architecture therefore matters more than the label. “Agent” should describe a system behavior: iterative model-driven decision making over tools, state and feedback — not merely a chatbot with a larger prompt.",{},{"id":271,"data":1621,"type":42,"tunes":1623},{"text":1622,"level":247},"The simplest example",{},{"id":276,"data":1625,"type":218,"tunes":1627},{"text":1626},"Suppose a developer asks an AI system: “Find why the test suite fails and fix the bug.” A single model call could only suggest likely causes from the text it was given.",{},{"id":281,"data":1629,"type":218,"tunes":1631},{"text":1630},"An agentic coding system can inspect the repository, search for the failing test, read relevant files, propose a change, edit the code, run the test, observe the failure, revise the implementation and run the test again.",{},{"id":286,"data":1633,"type":218,"tunes":1635},{"text":1634},"The agentic part is not simply that shell and file tools exist. It is that the model can use environmental feedback to choose the next step instead of following one completely predefined sequence.",{},{"id":291,"data":1637,"type":317,"tunes":1661},{"steps":1638,"title":1660,"orientation":316},[1639,1642,1645,1648,1651,1654,1657],{"label":1640,"description":1641},"1. Receive a goal","The user or upstream system defines the objective and relevant constraints.",{"label":1643,"description":1644},"2. Build current context","The runtime supplies instructions, state, history, memory, tools and current evidence.",{"label":1646,"description":1647},"3. Model decides next step","The model may answer, call a tool, request information, delegate or stop.",{"label":1649,"description":1650},"4. Runtime validates the request","Permissions, schemas, approvals and policy determine whether the proposed action may execute.",{"label":1652,"description":1653},"5. Execute tool or action","The external environment changes or returns new information.",{"label":1655,"description":1656},"6. Observe the result","The runtime feeds structured tool output, errors or state changes back into the next model step.",{"label":1658,"description":1659},"7. Continue or stop","The loop repeats until success, refusal, escalation, budget limit, timeout or another stopping condition.","The basic agent loop",{},{"id":320,"data":1663,"type":42,"tunes":1665},{"text":1664,"level":247},"Where the simple example stops",{},{"id":325,"data":1667,"type":218,"tunes":1669},{"text":1668},"Not every multi-step AI system is equally agentic. A workflow may use several LLM calls and tools while every step is predetermined in code. Another system may let the model decide which tool to call, in which order, how many times and when to stop.",{},{"id":330,"data":1671,"type":218,"tunes":1673},{"text":1672},"Both can be useful. The difference is where control lives. Predefined workflows put more control in application code. Agents move more tactical process decisions into the model\u002Fruntime loop.",{},{"id":335,"data":1675,"type":42,"tunes":1677},{"text":1676,"level":247},"Agent vs workflow",{},{"id":340,"data":1679,"type":369,"tunes":1698},{"rows":1680,"title":1693,"layout":361,"columns":1694},[1681,1684,1687,1690],{"id":344,"label":1682,"values":1683},"Process path",[347,347],{"id":349,"label":1685,"values":1686},"Tool sequence",[347,347],{"id":353,"label":1688,"values":1689},"Strength",[347,347],{"id":357,"label":1691,"values":1692},"Risk",[347,347],"Predefined workflow and agentic control",[1695,1697],{"id":364,"label":1696},"LLM workflow",{"id":367,"label":368},{},{"id":372,"data":1700,"type":218,"tunes":1702},{"text":1701},"Anthropic explicitly separates these two patterns: workflows orchestrate models and tools through predefined code paths, while agents let models dynamically direct their own processes and tool usage. This is not the only possible terminology, but it is a useful architecture boundary.",{},{"id":377,"data":1704,"type":42,"tunes":1706},{"text":1705,"level":247},"Agentic behavior is a spectrum, not a binary label",{},{"id":382,"data":1708,"type":361,"tunes":1738},{"content":1709,"stretched":43,"withHeadings":14},[1710,1714,1718,1722,1726,1730,1734],[1711,1712,1713],"Level","Example","Who decides the next step?",[1715,1716,1717],"Single model call","Summarize this document","Application calls model once",[1719,1720,1721],"Tool-assisted response","Model may use web search before answering","Model selects from bounded tools for one response",[1723,1724,1725],"Structured workflow","Classify → retrieve → generate → validate","Application workflow determines stages",[1727,1728,1729],"Adaptive workflow","Model can choose among several branches and retry","Shared control between application and model",[1731,1732,1733],"Agent loop","Model repeatedly chooses tools\u002Factions based on observations","Model directs tactical execution inside runtime constraints",[1735,1736,1737],"Long-running agent","Agent pauses, resumes, manages artifacts and continues","Model + persistent runtime manage evolving execution",{},{"id":415,"data":1740,"type":218,"tunes":1742},{"text":1741},"Calling every system above an “agent” can obscure important operational differences. The stronger the model's control over sequence, duration and actions, the more important runtime isolation, permissions, tracing, stopping conditions and trajectory evaluation become.",{},{"id":420,"data":1744,"type":42,"tunes":1746},{"text":1745,"level":247},"The minimum architecture of an agentic system",{},{"id":425,"data":1748,"type":361,"tunes":1788},{"content":1749,"stretched":43,"withHeadings":14},[1750,1753,1756,1758,1761,1764,1767,1770,1773,1776,1779,1782,1785],[1751,1752],"Component","Responsibility",[1754,1755],"Goal \u002F task","Defines what the system is trying to accomplish.",[435,1757],"Interprets context and decides the next action or output.",[1759,1760],"Instructions","Define role, constraints, priorities and task-specific policy.",[1762,1763],"Context assembler","Builds the information visible to the model on each step.",[1765,1766],"Tool catalog","Defines capabilities the model may request.",[1768,1769],"Runtime \u002F harness","Runs the loop, executes tools, manages state and handles stopping conditions.",[1771,1772],"Authorization layer","Determines whether a proposed action is permitted for the current principal.",[1774,1775],"State \u002F session","Preserves task progress across turns or execution steps.",[1777,1778],"Observation channel","Returns tool results and environment changes to the next model step.",[1780,1781],"Approvals \u002F human control","Pauses consequential actions where review is required.",[1783,1784],"Tracing \u002F audit","Records model calls, tools, transitions, approvals and failures.",[1786,1787],"Evaluation","Measures outcomes and execution trajectories against acceptance criteria.",{},{"id":469,"data":1790,"type":42,"tunes":1792},{"text":1791,"level":247},"A model is not an agent",{},{"id":474,"data":1794,"type":218,"tunes":1796},{"text":1795},"A language model produces outputs from inputs. It does not by itself own a filesystem, execute a shell command, maintain durable task state, enforce permissions or automatically call itself again.",{},{"id":479,"data":1798,"type":218,"tunes":1800},{"text":1799},"Those capabilities come from the surrounding runtime. The same model can behave as a simple chat model in one application and as the decision engine inside an agent loop in another.",{},{"id":484,"data":1802,"type":226,"tunes":1805},{"body":1803,"title":1804,"variant":488},"\u003Cstrong>Model capability determines what decisions can be proposed. Runtime architecture determines what can actually happen.\u003C\u002Fstrong>","Architecture rule",{},{"id":491,"data":1807,"type":42,"tunes":1809},{"text":1808,"level":247},"Tool use is central — but tool use alone does not make an agent",{},{"id":496,"data":1811,"type":218,"tunes":1813},{"text":1812},"Tools let the model acquire information and affect external systems. Examples include database reads, file operations, shell execution, web search, browser control, API calls, ticket updates or delegated specialist agents.",{},{"id":501,"data":1815,"type":218,"tunes":1817},{"text":1816},"A single model call can use one tool and still remain a bounded tool-assisted response rather than a long-running agent. Agentic behavior appears when tool observations feed an adaptive loop in which the model chooses what to do next.",{},{"id":506,"data":1819,"type":218,"tunes":1821},{"text":1820},"Tool design matters because tools are the contract between model reasoning and external reality. Ambiguous or overlapping tools create routing errors; large unstructured outputs pollute context; broad side-effect tools increase blast radius.",{},{"id":511,"data":1823,"type":42,"tunes":1825},{"text":1824,"level":247},"Tool capability, permission and authority are different",{},{"id":516,"data":1827,"type":361,"tunes":1853},{"content":1828,"stretched":43,"withHeadings":14},[1829,1832,1835,1838,1841,1844,1847,1850],[1830,1831],"Layer","Question",[1833,1834],"Capability","Can this runtime technically perform the operation?",[1836,1837],"Tool exposure","Is that capability available to this agent?",[1839,1840],"Permission","May this agent\u002Fsession use it under the current policy?",[1842,1843],"User authorization","Is the requesting principal allowed to cause this operation?",[1845,1846],"Business authority","Is the operation valid under domain rules, approvals and limits?",[1848,1849],"Execution","Did the operation actually occur?",[1851,1852],"Audit","Can the system prove who requested, approved and executed it?",{},{"id":545,"data":1855,"type":218,"tunes":1857},{"text":1856},"These layers are frequently collapsed in prototypes. A model sees a refund tool and therefore appears able to issue refunds. In production, the tool should still validate account, user, transaction, amount, policy and approval conditions independently of the model's request.",{},{"id":550,"data":1859,"type":42,"tunes":1861},{"text":1860,"level":247},"The runtime or harness is the actual execution system",{},{"id":555,"data":1863,"type":218,"tunes":1865},{"text":1864},"OpenAI's current agent documentation makes the runtime distinction explicit. Different runtimes can manage orchestration, state, tools, sandboxes and execution in different places, while the model remains only one part of the system.",{},{"id":560,"data":1867,"type":218,"tunes":1869},{"text":1868},"The Agents SDK describes a loop that repeatedly calls the current model, inspects the output, executes requested tools or handoffs, and continues until the model returns a final answer or another real stopping point.",{},{"id":565,"data":1871,"type":218,"tunes":1873},{"text":1872},"This means agent architecture decisions include where orchestration runs, where state lives, who executes tools, which sandbox contains side effects, and who owns retries, timeouts and resumability.",{},{"id":570,"data":1875,"type":42,"tunes":1877},{"text":1876,"level":247},"Planning is useful, but an explicit plan is not required",{},{"id":575,"data":1879,"type":218,"tunes":1881},{"text":1880},"Agents are often described as systems that “plan.” In practice, planning can be explicit or implicit. An agent may first produce a visible multi-step plan, or it may choose one next action at a time and revise after every observation.",{},{"id":580,"data":1883,"type":218,"tunes":1885},{"text":1884},"For highly uncertain tasks, short-horizon planning can be safer because the environment can invalidate a long plan. The architectural requirement is the ability to choose and revise actions based on the goal, current state and new evidence.",{},{"id":585,"data":1887,"type":42,"tunes":1889},{"text":1888,"level":247},"Environmental feedback is what makes the loop useful",{},{"id":590,"data":1891,"type":218,"tunes":1893},{"text":1892},"An agent becomes operationally meaningful when it can observe whether its action worked. Tool output, test results, API responses, filesystem state, browser state and application records provide external evidence that the system can use to revise its next decision.",{},{"id":595,"data":1895,"type":218,"tunes":1897},{"text":1896},"Anthropic's agent guidance emphasizes this feedback loop: agents use tools, obtain ground truth from the environment, assess progress and continue or request human input.",{},{"id":600,"data":1899,"type":226,"tunes":1902},{"body":1900,"title":1901,"variant":233},"An agent saying “the task is complete” does not prove completion. Where possible, verify the final state through an external system, test, file, transaction record or other observable outcome.","Self-report is not environmental proof",{},{"id":606,"data":1904,"type":42,"tunes":1906},{"text":1905,"level":247},"Agent state is not the same as model context",{},{"id":611,"data":1908,"type":218,"tunes":1910},{"text":1909},"A long-running task may need state that cannot or should not remain in the model context: task IDs, checkpoints, artifacts, approvals, external object identifiers, retry counters and workflow status.",{},{"id":616,"data":1912,"type":218,"tunes":1914},{"text":1913},"The runtime can preserve this durable state outside the model window and reconstruct the context required for the next step. This keeps model-visible context focused while maintaining continuity and resumability.",{},{"id":621,"data":1916,"type":42,"tunes":1918},{"text":1917,"level":247},"Memory is optional, not the definition of an agent",{},{"id":626,"data":1920,"type":218,"tunes":1922},{"text":1921},"An agent can operate successfully without long-term memory if the complete task fits inside one bounded run. Memory becomes useful when information must persist across sessions, tasks or long execution horizons.",{},{"id":631,"data":1924,"type":218,"tunes":1926},{"text":1925},"RAG, memory, state and context solve different problems. Treating a vector database as “the agent memory” or conversation history as “the state machine” usually hides important lifecycle and authority boundaries.",{},{"id":636,"data":1928,"type":642,"tunes":1933},{"url":1929,"title":1930,"excerpt":1931,"ctaLabel":1932},"https:\u002F\u002Fstajic.de\u002Fblog\u002Fai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context","AI Agent Memory Is Not RAG: How to Separate Memory, Retrieval, State and Context","A practical architecture separating persistent memory, authoritative application state, retrieval and the context supplied to the model.","Read the memory architecture article",{},{"id":645,"data":1935,"type":42,"tunes":1937},{"text":1936,"level":247},"Context engineering becomes dynamic in agents",{},{"id":650,"data":1939,"type":218,"tunes":1941},{"text":1940},"Every tool call can produce new context. Every step can also make earlier information obsolete. A strong agent runtime therefore rebuilds or curates context as execution progresses rather than replaying everything indefinitely.",{},{"id":655,"data":1943,"type":218,"tunes":1945},{"text":1944},"Tool definitions, task state, retrieved evidence, observations and memory all compete for the model's attention. Long-running agents need trimming, compaction or just-in-time loading so context remains relevant to the current decision.",{},{"id":660,"data":1947,"type":42,"tunes":1949},{"text":1948,"level":247},"Read tools and side-effect tools have different risk",{},{"id":665,"data":1951,"type":369,"tunes":1968},{"rows":1952,"title":1962,"layout":361,"columns":1963},[1953,1956,1959],{"id":669,"label":1954,"values":1955},"Examples",[347,347],{"id":673,"label":1957,"values":1958},"Main risk",[347,347],{"id":677,"label":1960,"values":1961},"Typical control",[347,347],"Information access versus external action",[1964,1966],{"id":683,"label":1965},"Read \u002F observe",{"id":686,"label":1967},"Write \u002F act",{},{"id":690,"data":1970,"type":42,"tunes":1972},{"text":1971,"level":247},"Human-in-the-loop is a control mechanism, not the opposite of agentic AI",{},{"id":695,"data":1974,"type":218,"tunes":1976},{"text":1975},"An agent does not stop being agentic because a human approves consequential steps. The model can still autonomously inspect, reason, search and prepare an action while the runtime requires human confirmation before execution.",{},{"id":700,"data":1978,"type":218,"tunes":1980},{"text":1979},"OpenAI's current agent safety guidance explicitly recommends approvals for tool operations in higher-risk workflows. Anthropic likewise emphasizes checkpoints and human judgment where agents encounter blockers or consequential decisions.",{},{"id":705,"data":1982,"type":218,"tunes":1984},{"text":1983},"The useful architecture question is not “human or autonomous?” but which decisions can be delegated, which require review and which must remain deterministic?",{},{"id":710,"data":1986,"type":42,"tunes":1988},{"text":1987,"level":247},"Agents need explicit stopping conditions",{},{"id":715,"data":1990,"type":361,"tunes":2022},{"content":1991,"stretched":43,"withHeadings":14},[1992,1995,1998,2001,2004,2007,2010,2013,2016,2019],[1993,1994],"Stopping condition","Purpose",[1996,1997],"Successful verified outcome","End when the external target state is confirmed.",[1999,2000],"Maximum steps","Prevent runaway loops.",[2002,2003],"Time budget","Bound wall-clock execution.",[2005,2006],"Cost\u002Ftoken budget","Limit resource consumption.",[2008,2009],"Repeated-action detector","Stop loops that are no longer making progress.",[2011,2012],"Permission boundary","Pause or stop when the next required action is not permitted.",[2014,2015],"Human approval checkpoint","Wait before consequential execution.",[2017,2018],"Unrecoverable tool failure","Escalate instead of retrying indefinitely.",[2020,2021],"Uncertainty threshold","Ask for clarification when the task cannot be safely inferred.",{},{"id":750,"data":2024,"type":42,"tunes":2026},{"text":2025,"level":247},"Recovery is part of agent behavior",{},{"id":755,"data":2028,"type":218,"tunes":2030},{"text":2029},"Agents operate in environments that fail: APIs time out, files change, credentials expire, webpages move and tools return malformed output. A useful agentic system therefore needs recovery behavior, not just a happy-path tool loop.",{},{"id":760,"data":2032,"type":218,"tunes":2034},{"text":2033},"Recovery can include retry with limits, choosing another tool, re-reading current state, asking the user, rolling back a partial action or escalating to a human.",{},{"id":765,"data":2036,"type":218,"tunes":2038},{"text":2037},"Retries also need idempotency awareness. Repeating a read is usually low risk; repeating a payment or message send can create duplicate side effects.",{},{"id":770,"data":2040,"type":42,"tunes":2042},{"text":2041,"level":247},"Agentic AI does not require multiple agents",{},{"id":775,"data":2044,"type":218,"tunes":2046},{"text":2045},"A single agent with a clear tool set is often simpler and easier to evaluate than a multi-agent architecture. Multiple agents are useful when specialization materially improves tool isolation, policy isolation, prompt clarity, ownership or trace legibility.",{},{"id":780,"data":2048,"type":218,"tunes":2050},{"text":2049},"OpenAI's current orchestration guidance explicitly recommends starting with one agent where possible and adding specialists only when the contract or ownership boundary materially changes.",{},{"id":785,"data":2052,"type":218,"tunes":2054},{"text":2053},"Multi-agent systems add new problems: delegation quality, duplicated context, conflicting state, handoff semantics, identity, cost and distributed failure handling.",{},{"id":790,"data":2056,"type":42,"tunes":2058},{"text":2057,"level":247},"Agent protocols are interoperability layers, not the agent itself",{},{"id":795,"data":2060,"type":218,"tunes":2062},{"text":2061},"Protocols such as MCP and A2A can make an agent architecture interoperable, but they do not create the agent loop by themselves. MCP can expose tools and resources. A2A can connect independently implemented agents. The application still needs runtime, authorization, state, evaluation and domain logic.",{},{"id":800,"data":2064,"type":218,"tunes":2066},{"text":2065},"This is why protocol capability must remain separate from business authority. Discovering a tool through MCP does not prove the current principal is allowed to use it. Receiving a task through A2A does not prove the remote agent may perform every requested action.",{},{"id":805,"data":2068,"type":642,"tunes":2072},{"url":807,"title":2069,"excerpt":2070,"ctaLabel":2071},"MCP vs A2A vs UCP vs AP2 vs A2UI: The Agent Protocol Stack Explained","A protocol-responsibility map showing why tool access, agent collaboration, commerce, payment authority and agent-driven UI belong to different interoperability boundaries.","Read the agent protocol stack",{},{"id":813,"data":2074,"type":42,"tunes":2076},{"text":2075,"level":247},"The trajectory is part of agent reliability",{},{"id":818,"data":2078,"type":218,"tunes":2080},{"text":2079},"A final answer is insufficient evidence for an agentic system because an agent can reach the right result through an unsafe or invalid path. It may use an unauthorized tool, skip a required check, retry a side effect, rely on stale state or accidentally succeed.",{},{"id":823,"data":2082,"type":218,"tunes":2084},{"text":2083},"Evaluation therefore needs execution traces: decisions, tool calls, approvals, observations, state changes and final outcome. Current OpenAI safety guidance recommends trace graders and evals; Anthropic's 2026 agent-evaluation guidance similarly treats multi-turn tool trajectories as first-class evaluation objects.",{},{"id":828,"data":2086,"type":218,"tunes":2088},{"text":2087},"The stronger reliability question is: Did the agent reach an acceptable outcome through an acceptable, recoverable and auditable trajectory?",{},{"id":833,"data":2090,"type":642,"tunes":2095},{"url":2091,"title":2092,"excerpt":2093,"ctaLabel":2094},"https:\u002F\u002Fstajic.de\u002Fblog\u002Fai-agent-reliability-why-the-final-answer-is-not-enough","AI Agent Reliability: Why the Final Answer Is Not Enough","Why production evaluation must inspect trajectories, tool use, state transitions and recoverability rather than only final answers.","Read the reliability article",{},{"id":841,"data":2097,"type":42,"tunes":2099},{"text":2098,"level":247},"Agentic systems increase the security surface",{},{"id":846,"data":2101,"type":361,"tunes":2138},{"content":2102,"stretched":43,"withHeadings":14},[2103,2106,2110,2114,2118,2122,2126,2130,2134],[1691,2104,2105],"Why agents amplify it","Architecture response",[2107,2108,2109],"Prompt injection","Untrusted content can influence future tool decisions","Separate instructions from data; constrain tools; sanitize or structure external input where possible",[2111,2112,2113],"Excessive permissions","Reasoning errors can become real side effects","Least privilege, scoped credentials, per-tool policy and approvals",[2115,2116,2117],"Credential exposure","Tools may need powerful secrets","Keep secrets outside model context; broker access through trusted runtime",[2119,2120,2121],"Confused deputy","Agent may act with authority broader than the requesting user","Bind execution to user\u002Fservice identity and re-authorize consequential actions",[2123,2124,2125],"Runaway loops","Model repeatedly calls tools without progress","Step, time and cost budgets plus loop detection",[2127,2128,2129],"State drift","Environment changes after the agent formed a plan","Re-read authoritative state before consequential actions",[2131,2132,2133],"Indirect injection","Tool\u002Fweb\u002Fdocument content contains instructions aimed at the model","Treat external content as untrusted data, not instruction authority",[2135,2136,2137],"Audit gap","Final result cannot show what was executed","Trace tool calls, approvals, identities and state changes",{},{"id":886,"data":2140,"type":42,"tunes":2142},{"text":2141,"level":247},"Agent observability must follow the loop",{},{"id":891,"data":2144,"type":218,"tunes":2146},{"text":2145},"Traditional service observability records requests, latency and errors. Agent observability needs an additional execution model: which agent was active, which model version made the decision, what context was available, which tool was selected, what arguments were sent, what result came back and why execution stopped.",{},{"id":896,"data":2148,"type":218,"tunes":2150},{"text":2149},"For sensitive systems, traces themselves require access control and retention policy because prompts, tool outputs and artifacts can contain confidential data.",{},{"id":901,"data":2152,"type":42,"tunes":2154},{"text":2153,"level":247},"How to evaluate an agentic system",{},{"id":906,"data":2156,"type":361,"tunes":2201},{"content":2157,"stretched":43,"withHeadings":14},[2158,2161,2165,2169,2173,2177,2181,2185,2189,2193,2197],[2159,1831,2160],"Dimension","Example evidence",[2162,2163,2164],"Task success","Did the requested outcome occur?","External state, tests, business outcome",[2166,2167,2168],"Trajectory quality","Were the steps acceptable?","Tool\u002Faction trace",[2170,2171,2172],"Tool selection","Did the agent choose appropriate capabilities?","Expected vs actual tool calls",[2174,2175,2176],"Permission adherence","Did it stay inside allowed authority?","Authorization logs and denied-action tests",[2178,2179,2180],"State handling","Did it use current authoritative state?","Freshness checks and state-change tests",[2182,2183,2184],"Recovery","Did it respond correctly to failures?","Injected timeout\u002Ferror scenarios",[2186,2187,2188],"Stopping behavior","Did it stop at the right point?","Step counts, loop detection, final-state proof",[2190,2191,2192],"Human escalation","Did it ask when review was required?","Approval\u002Fescalation traces",[2194,2195,2196],"Cost\u002Flatency","Was autonomy worth the operational cost?","Tokens, tool calls, duration",[2198,2199,2200],"Robustness","Does it survive realistic environment variation?","Repeated and adversarial trials",{},{"id":954,"data":2203,"type":42,"tunes":2205},{"text":2204,"level":247},"When an agent is appropriate",{},{"id":959,"data":2207,"type":361,"tunes":2233},{"content":2208,"stretched":43,"withHeadings":14},[2209,2212,2215,2218,2221,2224,2227,2230],[2210,2211],"Use an agent when","Prefer a workflow or simple call when",[2213,2214],"The number or order of steps cannot be known reliably in advance","The sequence is stable and deterministic",[2216,2217],"The system must inspect the environment and adapt","A single retrieval + generation step is sufficient",[2219,2220],"Several tools may be useful depending on intermediate results","One known API call solves the task",[2222,2223],"The task benefits from iterative verification or repair","The answer can be produced directly from supplied context",[2225,2226],"Failures require flexible recovery behavior","Failure branches are simple and can be encoded explicitly",[2228,2229],"Human review can be inserted at meaningful checkpoints","Every step is high-risk and must be manually controlled anyway",[2231,2232],"Expected value justifies extra latency, cost and complexity","Predictability and low cost matter more than flexibility",{},{"id":988,"data":2235,"type":218,"tunes":2237},{"text":2236},"A strong default is to start with the simplest solution that works and increase agentic complexity only when flexibility produces measurable value. Agents trade predictability, latency and cost for adaptive execution.",{},{"id":993,"data":2239,"type":42,"tunes":2241},{"text":2240,"level":247},"Original implementation evidence",{},{"id":998,"data":2243,"type":42,"tunes":2245},{"text":2244,"level":246},"Aaasaasa AI Client: model, runtime and permission are separate",{},{"id":1003,"data":2247,"type":218,"tunes":2249},{"text":2248},"Aaasaasa AI Client explicitly separates agent\u002Fclient, provider, model, runtime location and permissions. Its architecture documentation treats permissions as central tool\u002Fworkspace policy rather than a model property.",{},{"id":1008,"data":2251,"type":218,"tunes":2253},{"text":2252},"The same application can expose Direct Chat with no filesystem or shell tools while a Codex runtime operates under a selected workspace and permission profile. This demonstrates a core agentic architecture boundary: changing the runtime\u002Ftool surface changes what the system can do even when model access remains available.",{},{"id":1013,"data":2255,"type":218,"tunes":2257},{"text":2256},"The repository also distinguishes a local Codex runtime from model location: a local runtime can call a cloud model. This prevents the common mistake of equating “agent runs locally” with “inference is local.”",{},{"id":1018,"data":2259,"type":218,"tunes":2261},{"text":2260},"The implementation disables embedded execution paths whose approval semantics do not satisfy the required permission model. This supports the principle that agent capability should not bypass runtime authorization simply because an underlying framework can execute tools.",{},{"id":1023,"data":2263,"type":42,"tunes":2265},{"text":2264,"level":246},"Source of Truth Research Engine: bounded agentic research stages",{},{"id":1028,"data":2267,"type":218,"tunes":2269},{"text":2268},"The Source of Truth Research Engine uses a bounded research pipeline: discover → acquire → extract → verify → contradict → synthesize. Research jobs can execute through an AI runtime while evidence, sources, claims and contradictions remain in an external persistent store.",{},{"id":1033,"data":2271,"type":218,"tunes":2273},{"text":2272},"This is intentionally more controlled than an unconstrained autonomous research agent. The stages provide guardrails around what kind of work should happen next while still allowing model-driven research inside each bounded task.",{},{"id":1038,"data":2275,"type":218,"tunes":2277},{"text":2276},"That distinction is useful evidence for agent design: autonomy can be placed inside a structured delivery envelope rather than applied uniformly to the entire process.",{},{"id":1043,"data":2279,"type":361,"tunes":2302},{"content":2280,"stretched":43,"withHeadings":14},[2281,2284,2287,2290,2293,2296,2299],[2282,2283],"Implemented pattern","Agentic architecture lesson",[2285,2286],"Direct Chat has no OS tools","A model can exist without agentic execution capability.",[2288,2289],"Codex runtime has workspace permission profile","Tool authority belongs to runtime policy, not model capability.",[2291,2292],"Provider\u002Fmodel\u002Fruntime are separate concepts","Agent harness location and inference location are independent decisions.",[2294,2295],"Permission broker for tool-capable runtimes","Capability exposure can be centralized and governed.",[2297,2298],"Bounded research stages","Autonomy can operate inside explicit process boundaries.",[2300,2301],"Persistent claims\u002Fevidence outside model context","Agent state and evidence do not need to live only in conversation history.",{},{"id":1069,"data":2304,"type":226,"tunes":2307},{"body":2305,"title":2306,"variant":240},"These projects demonstrate concrete agent\u002Fruntime, permission and bounded-research patterns. They are not presented as proof of large-scale commercial autonomous-agent deployment.","Evidence boundary",{},{"id":1075,"data":2309,"type":42,"tunes":2311},{"text":2310,"level":247},"Common agentic AI failure modes",{},{"id":1080,"data":2313,"type":361,"tunes":2354},{"content":2314,"stretched":43,"withHeadings":14},[2315,2318,2321,2324,2327,2330,2333,2336,2339,2342,2345,2348,2351],[2316,2317],"Failure mode","What actually failed",[2319,2320],"“Agent” is only a chatbot with tools listed in the prompt","No reliable runtime loop or tool execution architecture exists",[2322,2323],"Tool support is treated as permission","Capability and authorization boundaries are collapsed",[2325,2326],"Agent trusts its own completion statement","Outcome is not verified against external state",[2328,2329],"Every task becomes multi-agent","Complexity increases without a real ownership or specialization boundary",[2331,2332],"Conversation history is used as durable state","Resumability and authoritative state become fragile",[2334,2335],"Agent retries side effects blindly","Duplicate messages, payments or state changes become possible",[2337,2338],"No step\u002Fcost limits","Agent can loop indefinitely or consume uncontrolled resources",[2340,2341],"Tool output is trusted as instruction","Indirect prompt injection can redirect behavior",[2343,2344],"Correct final answer is the only evaluation","Unsafe or invalid trajectories remain invisible",[2346,2347],"Model upgrade is treated as transparent","Tool selection, planning and stopping behavior can change",[2349,2350],"One broad tool exposes many privileged operations","Blast radius increases and intent becomes harder to validate",[2352,2353],"Human approval exists but reviewer lacks context","Approval becomes ceremonial rather than effective",{},{"id":1124,"data":2356,"type":42,"tunes":2358},{"text":2357,"level":247},"Common misconceptions",{},{"id":1129,"data":2360,"type":361,"tunes":2395},{"content":2361,"stretched":43,"withHeadings":14},[2362,2365,2368,2371,2374,2377,2380,2383,2386,2389,2392],[2363,2364],"Misconception","Correction",[2366,2367],"“An LLM is an agent.”","The model is the decision component; the agent is the surrounding system that manages tools, state and iteration.",[2369,2370],"“Tool calling automatically means agentic AI.”","A single bounded tool call may not involve an adaptive multi-step agent loop.",[2372,2373],"“Agents must be fully autonomous.”","Agentic systems can require approvals and operate under narrow permission boundaries.",[2375,2376],"“Agents need long-term memory.”","Memory is optional; many useful agents complete bounded tasks without cross-session memory.",[2378,2379],"“Agents must create a written plan first.”","Planning can be explicit or implicit and can occur one step at a time.",[2381,2382],"“Multi-agent is more advanced than single-agent.”","It is more complex; use it only when specialization or ownership boundaries justify it.",[2384,2385],"“MCP creates an agent.”","MCP exposes tools\u002Fresources; the runtime still needs an agent loop and authorization model.",[2387,2388],"“A local runtime means the model is local.”","Runtime location and inference\u002Fprovider location are separate.",[2390,2391],"“If the final result is correct, the agent worked correctly.”","An unsafe or unauthorized trajectory can still produce a correct result.",[2393,2394],"“Human approval removes autonomy.”","Approval can constrain selected actions while the rest of the process remains model-directed.",{},{"id":1167,"data":2397,"type":42,"tunes":2399},{"text":2398,"level":247},"A practical agent design sequence",{},{"id":1172,"data":2401,"type":317,"tunes":2440},{"steps":2402,"title":2439,"orientation":316},[2403,2406,2409,2412,2415,2418,2421,2424,2427,2430,2433,2436],{"label":2404,"description":2405},"1. Define the outcome","State what external result or artifact proves task success.",{"label":2407,"description":2408},"2. Decide whether an agent is actually needed","Prefer a simple call or deterministic workflow when the path is predictable.",{"label":2410,"description":2411},"3. Identify state and Source of Truth","Define which systems own current facts, task progress and business state.",{"label":2413,"description":2414},"4. Define the tool surface","Expose the smallest set of clear capabilities required for the task.",{"label":2416,"description":2417},"5. Bind identity and permissions","Separate user authority, agent\u002Fruntime permissions and tool capabilities.",{"label":2419,"description":2420},"6. Choose autonomy boundaries","Specify what the model may decide dynamically and what remains deterministic.",{"label":2422,"description":2423},"7. Add approval checkpoints","Require review before consequential or irreversible actions where appropriate.",{"label":2425,"description":2426},"8. Define stopping and recovery","Set success proof, budgets, timeouts, retries, escalation and loop controls.",{"label":2428,"description":2429},"9. Design context\u002Fstate management","Keep current state, memory, tool observations and durable artifacts in the correct layers.",{"label":2431,"description":2432},"10. Trace the trajectory","Record enough execution structure to debug and audit model\u002Ftool decisions.",{"label":2434,"description":2435},"11. Evaluate realistic failures","Test stale state, tool errors, prompt injection, ambiguous requests and changed environments.",{"label":2437,"description":2438},"12. Expand autonomy only from evidence","Increase permissions or execution horizon when evaluation shows the benefit justifies the risk.","Design the agent from authority outward",{},{"id":1214,"data":2442,"type":42,"tunes":2444},{"text":2443,"level":247},"Agentic AI architecture checklist",{},{"id":1219,"data":2446,"type":361,"tunes":2489},{"content":2447,"stretched":43,"withHeadings":14},[2448,2450,2453,2456,2459,2462,2465,2468,2471,2474,2477,2480,2483,2486],[1831,2449],"Expected evidence",[2451,2452],"What proves success?","External outcome, artifact, test or authoritative state.",[2454,2455],"Why is an agent needed?","The path genuinely depends on intermediate observations.",[2457,2458],"Which decisions are model-driven?","Explicit autonomy boundary.",[2460,2461],"Which tools exist?","Small, documented, unambiguous capability set.",[2463,2464],"Who may use each tool?","Identity- and context-aware authorization policy.",[2466,2467],"Which actions need approval?","Consequence-based review rules.",[2469,2470],"Where does task state live?","Application-owned state separate from transient model context.",[2472,2473],"How does the agent recover?","Retry, re-read, rollback, clarification and escalation behavior.",[2475,2476],"How does it stop?","Verified completion plus step\u002Ftime\u002Fcost limits.",[2478,2479],"How are side effects protected?","Validation, idempotency, least privilege and confirmation.",[2481,2482],"Can execution be reconstructed?","Tool, approval and state-transition traces.",[2484,2485],"How is it evaluated?","Outcome + trajectory + robustness tests.",[2487,2488],"What changes after a model\u002Fruntime update?","Regression suite for tool selection, permissions, stopping and recovery.",{},{"id":1265,"data":2491,"type":42,"tunes":2493},{"text":2492,"level":247},"Edge cases and limitations",{},{"id":1270,"data":2495,"type":218,"tunes":2497},{"text":2496},"Some systems are “agentic” only in a narrow routing sense: the model selects one specialist or tool and then the rest of the workflow is deterministic. That can still be useful, but it should not be described as equivalent to a long-running autonomous agent.",{},{"id":1275,"data":2499,"type":218,"tunes":2501},{"text":2500},"Highly consequential domains may intentionally restrict agent autonomy. An AI system can inspect evidence, prepare recommendations and fill structured forms while a human remains the only actor allowed to commit the final transaction.",{},{"id":1280,"data":2503,"type":218,"tunes":2505},{"text":2504},"Some environments are well suited to agents because feedback is objective. Coding agents can run tests; infrastructure agents can inspect metrics; data agents can validate query results. Open-ended domains with weak feedback require more cautious evaluation.",{},{"id":1285,"data":2507,"type":218,"tunes":2509},{"text":2508},"An agent can operate entirely locally, entirely through managed cloud services or in a hybrid architecture. Agentic behavior describes control flow, not hosting location.",{},{"id":1290,"data":2511,"type":218,"tunes":2513},{"text":2512},"The term “reasoning” should not be used as proof that the agent's internal process is correct. Production assurance should rely on observable inputs, actions, outputs, state and evaluation rather than unverifiable claims about hidden reasoning.",{},{"id":1295,"data":2515,"type":42,"tunes":2517},{"text":2516,"level":247},"What would change this answer?",{},{"id":1300,"data":2519,"type":218,"tunes":2521},{"text":2520},"Vendor APIs and agent frameworks will continue to evolve, but the architecture boundary is stable: a model proposes decisions, a runtime manages the loop, tools connect to the environment, permissions constrain actions and external observations determine what actually happened.",{},{"id":1305,"data":2523,"type":218,"tunes":2525},{"text":2524},"As models become more reliable, systems may safely delegate longer horizons or more complex recovery behavior. As runtime verification and authorization improve, some approval steps may become automated. Those are changes in autonomy level, not changes to the fundamental responsibility layers.",{},{"id":1310,"data":2527,"type":218,"tunes":2529},{"text":2528},"The recommended architecture also changes by consequence. A research agent that only reads public sources can tolerate different controls from an agent that writes production configuration or moves money.",{},{"id":1315,"data":2531,"type":42,"tunes":2533},{"text":2532,"level":247},"Related canonical knowledge",{},{"id":1320,"data":2535,"type":218,"tunes":2537},{"text":2536},"Agentic AI sits above several prerequisite layers: context engineering determines what the model sees; Source-of-Truth architecture determines which information is authoritative; retrieval supplies external evidence; runtime architecture determines what can execute.",{},{"id":1325,"data":2539,"type":218,"tunes":2541},{"text":2540},"Downstream nodes include tool calling, MCP, A2A, agent identity, permissions, auditability, human-in-the-loop, orchestration, memory and multi-agent systems.",{},{"id":1330,"data":2543,"type":218,"tunes":2545},{"text":2544},"The protocol stack article should therefore be read after the basic agent concept: protocols standardize boundaries around agents; they do not define agentic behavior itself.",{},{"id":1335,"data":2547,"type":42,"tunes":2549},{"text":2548,"level":247},"Frequently asked questions",{},{"id":1340,"data":2551,"type":1340,"tunes":2581},{"items":2552,"title":2580},[2553,2556,2559,2562,2565,2568,2571,2574,2577],{"id":1344,"answer":2554,"question":2555},"Agentic AI is an AI system in which a model can pursue a goal over multiple steps by choosing actions or tools, observing results, updating its state and continuing until a stopping condition is reached.","What is agentic AI?",{"id":1348,"answer":2557,"question":2558},"An LLM produces outputs from inputs. An agent combines a model with a runtime, tools, state, permissions, context management and an iterative execution loop.","What is the difference between an LLM and an AI agent?",{"id":1352,"answer":2560,"question":2561},"Not necessarily. A single tool-assisted model response can be bounded and non-agentic. Agentic behavior appears when tool observations drive an adaptive multi-step loop.","Does tool calling make a system an agent?",{"id":1356,"answer":2563,"question":2564},"A workflow usually follows a process path defined in application code. An agent has more model-driven control over which steps and tools to use based on intermediate observations.","What is the difference between an agent and an AI workflow?",{"id":1360,"answer":2566,"question":2567},"No. Long-term memory is useful for persistent information across sessions, but many agents complete bounded tasks using only current task state and context.","Do agents need memory?",{"id":1364,"answer":2569,"question":2570},"No. A single agent is often simpler. Multi-agent systems are justified when specialization, tool isolation, policy isolation or ownership boundaries materially improve the system.","Do AI agents need multiple agents?",{"id":1368,"answer":2572,"question":2573},"Yes. The agent can autonomously perform low-risk analysis and preparation while the runtime pauses for human approval before consequential actions.","Can an agent be human-in-the-loop?",{"id":1372,"answer":2575,"question":2576},"No. MCP is an interoperability protocol for exposing tools, resources and prompts. An agent runtime can use MCP, but still needs its own loop, state, authorization and evaluation.","Is MCP an agent framework?",{"id":1376,"answer":2578,"question":2579},"Where possible, verify success through external state, tests, artifacts or authoritative system records rather than trusting the model's own completion statement.","How do you know an agent actually completed a task?","Agentic AI FAQ",{},{"id":1382,"data":2583,"type":42,"tunes":2585},{"text":2584,"level":247},"Glossary",{},{"id":1387,"data":2587,"type":1387,"tunes":2620},{"title":2588,"entries":2589},"Key agentic AI terms",[2590,2593,2595,2597,2599,2602,2605,2608,2611,2614,2617],{"term":2591,"anchor":1393,"definition":2592},"Agentic AI","AI system behavior in which a model dynamically directs multi-step execution using tools, observations and state toward a goal.",{"term":1396,"anchor":1397,"definition":2594},"A model-centered system with runtime, tools, state and an execution loop that can pursue a task over multiple steps.",{"term":1731,"anchor":1401,"definition":2596},"Repeated cycle of model decision, tool\u002Faction execution, observation and updated model decision until stopping.",{"term":1768,"anchor":1404,"definition":2598},"The execution layer that manages the model loop, tools, state, approvals, context, errors and stopping conditions.",{"term":2600,"anchor":1408,"definition":2601},"Tool","A capability exposed to the model for reading information, computing, delegating or changing external state.",{"term":2603,"anchor":1412,"definition":2604},"Observation","Information returned from a tool or environment and supplied to a later agent step.",{"term":2606,"anchor":1416,"definition":2607},"Agent state","Persistent task or execution information that exists outside a single model output and may survive across steps or pauses.",{"term":2609,"anchor":1420,"definition":2610},"Autonomy boundary","The explicit limit defining which decisions and actions the model may control dynamically.",{"term":2612,"anchor":1424,"definition":2613},"Human-in-the-loop","A control pattern in which human review, input or approval is required at selected points in an AI-driven process.",{"term":2615,"anchor":1428,"definition":2616},"Trajectory","The sequence of relevant states, decisions, tool calls, actions and observations between task request and final outcome.",{"term":2618,"anchor":1432,"definition":2619},"Idempotency","Property that allows an operation to be repeated without unintentionally applying the same side effect multiple times.",{},{"id":1436,"data":2622,"type":42,"tunes":2624},{"text":2623,"level":247},"Conclusion",{},{"id":1441,"data":2626,"type":218,"tunes":2628},{"text":2627},"Agentic AI is not simply a smarter model or a chatbot with more tools. It is a system architecture in which a model participates in an iterative control loop: decide, act, observe, update and continue.",{},{"id":1446,"data":2630,"type":218,"tunes":2632},{"text":2631},"The model provides flexible decision making, but the surrounding runtime must own execution reality: permissions, tool access, state, approvals, retries, budgets, stopping conditions, tracing and verification.",{},{"id":1451,"data":2634,"type":218,"tunes":2636},{"text":2635},"The most useful design principle is therefore: delegate tactical choice to the model only inside explicit technical and business boundaries. Agentic capability becomes production capability only when autonomy, authority and evidence remain separable.",{},{"id":1456,"data":2638,"type":42,"tunes":2640},{"text":2639,"level":247},"Primary sources and current guidance",{},{"id":1461,"data":2642,"type":218,"tunes":2644},{"text":2643},"The sources below support the current architectural distinctions around agents, workflows, loops, tools, orchestration, safety and evaluation. Project sections are original implementation evidence and are explicitly bounded to what the repositories demonstrate.",{},{"id":1466,"data":2646,"type":1473,"tunes":2651},{"link":1468,"meta":2647},{"image":2648,"title":2649,"description":2650},{"url":347},"OpenAI — Agents","Current developer guidance defining runtime choices for multi-step work, tools, state, orchestration and agent execution.",{},{"id":1476,"data":2653,"type":1473,"tunes":2658},{"link":1478,"meta":2654},{"image":2655,"title":2656,"description":2657},{"url":347},"OpenAI — Agent definitions","Current documentation describing an agent as a model plus instructions and optional runtime behavior including tools, guardrails, MCP servers and handoffs.",{},{"id":1485,"data":2660,"type":1473,"tunes":2665},{"link":1487,"meta":2661},{"image":2662,"title":2663,"description":2664},{"url":347},"OpenAI — Running agents","Current documentation of the agent loop: model call, tool execution or handoff, continuation and final stopping point.",{},{"id":1494,"data":2667,"type":1473,"tunes":2672},{"link":1496,"meta":2668},{"image":2669,"title":2670,"description":2671},{"url":347},"OpenAI — Orchestration and handoffs","Current guidance on handoffs, agents-as-tools and when specialist agents add useful ownership or capability boundaries.",{},{"id":1503,"data":2674,"type":1473,"tunes":2679},{"link":1505,"meta":2675},{"image":2676,"title":2677,"description":2678},{"url":347},"OpenAI — Safety in building agents","Current safety guidance covering tool approvals, prompt injection, guardrails and trace-based evaluation.",{},{"id":1512,"data":2681,"type":1473,"tunes":2686},{"link":1514,"meta":2682},{"image":2683,"title":2684,"description":2685},{"url":347},"Anthropic — Building effective agents","Engineering guidance distinguishing predefined workflows from model-directed agents and describing tool-based environmental feedback loops.",{},{"id":1521,"data":2688,"type":1473,"tunes":2693},{"link":1523,"meta":2689},{"image":2690,"title":2691,"description":2692},{"url":347},"Anthropic — Effective context engineering for AI agents","Practical framing of agents as LLMs autonomously using tools in a loop, with dynamic just-in-time context management.",{},{"id":1530,"data":2695,"type":1473,"tunes":2700},{"link":1532,"meta":2696},{"image":2697,"title":2698,"description":2699},{"url":347},"Anthropic — Demystifying evals for AI agents","2026 guidance on evaluating multi-turn agents that call tools, modify state and adapt to intermediate results.",{},"2.31.6","Agentic AI uses models inside multi-step execution loops where they can choose tools, observe results, update state and adapt their next action within explicit runtime and permission boundaries.",{"lang":7,"title":208,"content":210,"contentJson":2704,"excerpt":1539},{"time":212,"blocks":2705,"version":1538},[2706,2709,2712,2715,2718,2721,2724,2727,2730,2733,2736,2739,2742,2745,2756,2759,2762,2765,2768,2783,2786,2789,2800,2803,2806,2823,2826,2829,2832,2835,2838,2841,2844,2847,2850,2862,2865,2868,2871,2874,2877,2880,2883,2886,2889,2892,2895,2898,2901,2904,2907,2910,2913,2916,2919,2922,2925,2928,2931,2944,2947,2950,2953,2956,2959,2973,2976,2979,2982,2985,2988,2991,2994,2997,3000,3003,3006,3009,3012,3015,3018,3021,3024,3027,3040,3043,3046,3049,3052,3067,3070,3082,3085,3088,3091,3094,3097,3100,3103,3106,3109,3112,3115,3126,3129,3132,3149,3152,3167,3170,3186,3189,3207,3210,3213,3216,3219,3222,3225,3228,3231,3234,3237,3240,3243,3246,3249,3252,3265,3268,3283,3286,3289,3292,3295,3298,3301,3306,3311,3316,3321,3326,3331,3336],{"id":215,"data":2707,"type":218,"tunes":2708},{"text":217},{},{"id":221,"data":2710,"type":226,"tunes":2711},{"body":223,"title":224,"variant":225},{},{"id":229,"data":2713,"type":226,"tunes":2714},{"body":231,"title":232,"variant":233},{},{"id":236,"data":2716,"type":226,"tunes":2717},{"body":238,"title":239,"variant":240},{},{"id":243,"data":2719,"type":248,"tunes":2720},{"title":245,"maxLevel":246,"minLevel":247},{},{"id":251,"data":2722,"type":42,"tunes":2723},{"text":253,"level":247},{},{"id":256,"data":2725,"type":218,"tunes":2726},{"text":258},{},{"id":261,"data":2728,"type":218,"tunes":2729},{"text":263},{},{"id":266,"data":2731,"type":218,"tunes":2732},{"text":268},{},{"id":271,"data":2734,"type":42,"tunes":2735},{"text":273,"level":247},{},{"id":276,"data":2737,"type":218,"tunes":2738},{"text":278},{},{"id":281,"data":2740,"type":218,"tunes":2741},{"text":283},{},{"id":286,"data":2743,"type":218,"tunes":2744},{"text":288},{},{"id":291,"data":2746,"type":317,"tunes":2755},{"steps":2747,"title":315,"orientation":316},[2748,2749,2750,2751,2752,2753,2754],{"label":295,"description":296},{"label":298,"description":299},{"label":301,"description":302},{"label":304,"description":305},{"label":307,"description":308},{"label":310,"description":311},{"label":313,"description":314},{},{"id":320,"data":2757,"type":42,"tunes":2758},{"text":322,"level":247},{},{"id":325,"data":2760,"type":218,"tunes":2761},{"text":327},{},{"id":330,"data":2763,"type":218,"tunes":2764},{"text":332},{},{"id":335,"data":2766,"type":42,"tunes":2767},{"text":337,"level":247},{},{"id":340,"data":2769,"type":369,"tunes":2782},{"rows":2770,"title":360,"layout":361,"columns":2779},[2771,2773,2775,2777],{"id":344,"label":345,"values":2772},[347,347],{"id":349,"label":350,"values":2774},[347,347],{"id":353,"label":354,"values":2776},[347,347],{"id":357,"label":358,"values":2778},[347,347],[2780,2781],{"id":364,"label":365},{"id":367,"label":368},{},{"id":372,"data":2784,"type":218,"tunes":2785},{"text":374},{},{"id":377,"data":2787,"type":42,"tunes":2788},{"text":379,"level":247},{},{"id":382,"data":2790,"type":361,"tunes":2799},{"content":2791,"stretched":43,"withHeadings":14},[2792,2793,2794,2795,2796,2797,2798],[386,387,388],[390,391,392],[394,395,396],[398,399,400],[402,403,404],[406,407,408],[410,411,412],{},{"id":415,"data":2801,"type":218,"tunes":2802},{"text":417},{},{"id":420,"data":2804,"type":42,"tunes":2805},{"text":422,"level":247},{},{"id":425,"data":2807,"type":361,"tunes":2822},{"content":2808,"stretched":43,"withHeadings":14},[2809,2810,2811,2812,2813,2814,2815,2816,2817,2818,2819,2820,2821],[429,430],[432,433],[435,436],[438,439],[441,442],[444,445],[447,448],[450,451],[453,454],[456,457],[459,460],[462,463],[465,466],{},{"id":469,"data":2824,"type":42,"tunes":2825},{"text":471,"level":247},{},{"id":474,"data":2827,"type":218,"tunes":2828},{"text":476},{},{"id":479,"data":2830,"type":218,"tunes":2831},{"text":481},{},{"id":484,"data":2833,"type":226,"tunes":2834},{"body":486,"title":487,"variant":488},{},{"id":491,"data":2836,"type":42,"tunes":2837},{"text":493,"level":247},{},{"id":496,"data":2839,"type":218,"tunes":2840},{"text":498},{},{"id":501,"data":2842,"type":218,"tunes":2843},{"text":503},{},{"id":506,"data":2845,"type":218,"tunes":2846},{"text":508},{},{"id":511,"data":2848,"type":42,"tunes":2849},{"text":513,"level":247},{},{"id":516,"data":2851,"type":361,"tunes":2861},{"content":2852,"stretched":43,"withHeadings":14},[2853,2854,2855,2856,2857,2858,2859,2860],[520,521],[523,524],[526,527],[529,530],[532,533],[535,536],[538,539],[541,542],{},{"id":545,"data":2863,"type":218,"tunes":2864},{"text":547},{},{"id":550,"data":2866,"type":42,"tunes":2867},{"text":552,"level":247},{},{"id":555,"data":2869,"type":218,"tunes":2870},{"text":557},{},{"id":560,"data":2872,"type":218,"tunes":2873},{"text":562},{},{"id":565,"data":2875,"type":218,"tunes":2876},{"text":567},{},{"id":570,"data":2878,"type":42,"tunes":2879},{"text":572,"level":247},{},{"id":575,"data":2881,"type":218,"tunes":2882},{"text":577},{},{"id":580,"data":2884,"type":218,"tunes":2885},{"text":582},{},{"id":585,"data":2887,"type":42,"tunes":2888},{"text":587,"level":247},{},{"id":590,"data":2890,"type":218,"tunes":2891},{"text":592},{},{"id":595,"data":2893,"type":218,"tunes":2894},{"text":597},{},{"id":600,"data":2896,"type":226,"tunes":2897},{"body":602,"title":603,"variant":233},{},{"id":606,"data":2899,"type":42,"tunes":2900},{"text":608,"level":247},{},{"id":611,"data":2902,"type":218,"tunes":2903},{"text":613},{},{"id":616,"data":2905,"type":218,"tunes":2906},{"text":618},{},{"id":621,"data":2908,"type":42,"tunes":2909},{"text":623,"level":247},{},{"id":626,"data":2911,"type":218,"tunes":2912},{"text":628},{},{"id":631,"data":2914,"type":218,"tunes":2915},{"text":633},{},{"id":636,"data":2917,"type":642,"tunes":2918},{"url":638,"title":639,"excerpt":640,"ctaLabel":641},{},{"id":645,"data":2920,"type":42,"tunes":2921},{"text":647,"level":247},{},{"id":650,"data":2923,"type":218,"tunes":2924},{"text":652},{},{"id":655,"data":2926,"type":218,"tunes":2927},{"text":657},{},{"id":660,"data":2929,"type":42,"tunes":2930},{"text":662,"level":247},{},{"id":665,"data":2932,"type":369,"tunes":2943},{"rows":2933,"title":680,"layout":361,"columns":2940},[2934,2936,2938],{"id":669,"label":670,"values":2935},[347,347],{"id":673,"label":674,"values":2937},[347,347],{"id":677,"label":678,"values":2939},[347,347],[2941,2942],{"id":683,"label":684},{"id":686,"label":687},{},{"id":690,"data":2945,"type":42,"tunes":2946},{"text":692,"level":247},{},{"id":695,"data":2948,"type":218,"tunes":2949},{"text":697},{},{"id":700,"data":2951,"type":218,"tunes":2952},{"text":702},{},{"id":705,"data":2954,"type":218,"tunes":2955},{"text":707},{},{"id":710,"data":2957,"type":42,"tunes":2958},{"text":712,"level":247},{},{"id":715,"data":2960,"type":361,"tunes":2972},{"content":2961,"stretched":43,"withHeadings":14},[2962,2963,2964,2965,2966,2967,2968,2969,2970,2971],[719,720],[722,723],[725,726],[728,729],[731,732],[734,735],[737,738],[740,741],[743,744],[746,747],{},{"id":750,"data":2974,"type":42,"tunes":2975},{"text":752,"level":247},{},{"id":755,"data":2977,"type":218,"tunes":2978},{"text":757},{},{"id":760,"data":2980,"type":218,"tunes":2981},{"text":762},{},{"id":765,"data":2983,"type":218,"tunes":2984},{"text":767},{},{"id":770,"data":2986,"type":42,"tunes":2987},{"text":772,"level":247},{},{"id":775,"data":2989,"type":218,"tunes":2990},{"text":777},{},{"id":780,"data":2992,"type":218,"tunes":2993},{"text":782},{},{"id":785,"data":2995,"type":218,"tunes":2996},{"text":787},{},{"id":790,"data":2998,"type":42,"tunes":2999},{"text":792,"level":247},{},{"id":795,"data":3001,"type":218,"tunes":3002},{"text":797},{},{"id":800,"data":3004,"type":218,"tunes":3005},{"text":802},{},{"id":805,"data":3007,"type":642,"tunes":3008},{"url":807,"title":808,"excerpt":809,"ctaLabel":810},{},{"id":813,"data":3010,"type":42,"tunes":3011},{"text":815,"level":247},{},{"id":818,"data":3013,"type":218,"tunes":3014},{"text":820},{},{"id":823,"data":3016,"type":218,"tunes":3017},{"text":825},{},{"id":828,"data":3019,"type":218,"tunes":3020},{"text":830},{},{"id":833,"data":3022,"type":642,"tunes":3023},{"url":835,"title":836,"excerpt":837,"ctaLabel":838},{},{"id":841,"data":3025,"type":42,"tunes":3026},{"text":843,"level":247},{},{"id":846,"data":3028,"type":361,"tunes":3039},{"content":3029,"stretched":43,"withHeadings":14},[3030,3031,3032,3033,3034,3035,3036,3037,3038],[358,850,851],[853,854,855],[857,858,859],[861,862,863],[865,866,867],[869,870,871],[873,874,875],[877,878,879],[881,882,883],{},{"id":886,"data":3041,"type":42,"tunes":3042},{"text":888,"level":247},{},{"id":891,"data":3044,"type":218,"tunes":3045},{"text":893},{},{"id":896,"data":3047,"type":218,"tunes":3048},{"text":898},{},{"id":901,"data":3050,"type":42,"tunes":3051},{"text":903,"level":247},{},{"id":906,"data":3053,"type":361,"tunes":3066},{"content":3054,"stretched":43,"withHeadings":14},[3055,3056,3057,3058,3059,3060,3061,3062,3063,3064,3065],[910,521,911],[913,914,915],[917,918,919],[921,922,923],[925,926,927],[929,930,931],[933,934,935],[937,938,939],[941,942,943],[945,946,947],[949,950,951],{},{"id":954,"data":3068,"type":42,"tunes":3069},{"text":956,"level":247},{},{"id":959,"data":3071,"type":361,"tunes":3081},{"content":3072,"stretched":43,"withHeadings":14},[3073,3074,3075,3076,3077,3078,3079,3080],[963,964],[966,967],[969,970],[972,973],[975,976],[978,979],[981,982],[984,985],{},{"id":988,"data":3083,"type":218,"tunes":3084},{"text":990},{},{"id":993,"data":3086,"type":42,"tunes":3087},{"text":995,"level":247},{},{"id":998,"data":3089,"type":42,"tunes":3090},{"text":1000,"level":246},{},{"id":1003,"data":3092,"type":218,"tunes":3093},{"text":1005},{},{"id":1008,"data":3095,"type":218,"tunes":3096},{"text":1010},{},{"id":1013,"data":3098,"type":218,"tunes":3099},{"text":1015},{},{"id":1018,"data":3101,"type":218,"tunes":3102},{"text":1020},{},{"id":1023,"data":3104,"type":42,"tunes":3105},{"text":1025,"level":246},{},{"id":1028,"data":3107,"type":218,"tunes":3108},{"text":1030},{},{"id":1033,"data":3110,"type":218,"tunes":3111},{"text":1035},{},{"id":1038,"data":3113,"type":218,"tunes":3114},{"text":1040},{},{"id":1043,"data":3116,"type":361,"tunes":3125},{"content":3117,"stretched":43,"withHeadings":14},[3118,3119,3120,3121,3122,3123,3124],[1047,1048],[1050,1051],[1053,1054],[1056,1057],[1059,1060],[1062,1063],[1065,1066],{},{"id":1069,"data":3127,"type":226,"tunes":3128},{"body":1071,"title":1072,"variant":240},{},{"id":1075,"data":3130,"type":42,"tunes":3131},{"text":1077,"level":247},{},{"id":1080,"data":3133,"type":361,"tunes":3148},{"content":3134,"stretched":43,"withHeadings":14},[3135,3136,3137,3138,3139,3140,3141,3142,3143,3144,3145,3146,3147],[1084,1085],[1087,1088],[1090,1091],[1093,1094],[1096,1097],[1099,1100],[1102,1103],[1105,1106],[1108,1109],[1111,1112],[1114,1115],[1117,1118],[1120,1121],{},{"id":1124,"data":3150,"type":42,"tunes":3151},{"text":1126,"level":247},{},{"id":1129,"data":3153,"type":361,"tunes":3166},{"content":3154,"stretched":43,"withHeadings":14},[3155,3156,3157,3158,3159,3160,3161,3162,3163,3164,3165],[1133,1134],[1136,1137],[1139,1140],[1142,1143],[1145,1146],[1148,1149],[1151,1152],[1154,1155],[1157,1158],[1160,1161],[1163,1164],{},{"id":1167,"data":3168,"type":42,"tunes":3169},{"text":1169,"level":247},{},{"id":1172,"data":3171,"type":317,"tunes":3185},{"steps":3172,"title":1211,"orientation":316},[3173,3174,3175,3176,3177,3178,3179,3180,3181,3182,3183,3184],{"label":1176,"description":1177},{"label":1179,"description":1180},{"label":1182,"description":1183},{"label":1185,"description":1186},{"label":1188,"description":1189},{"label":1191,"description":1192},{"label":1194,"description":1195},{"label":1197,"description":1198},{"label":1200,"description":1201},{"label":1203,"description":1204},{"label":1206,"description":1207},{"label":1209,"description":1210},{},{"id":1214,"data":3187,"type":42,"tunes":3188},{"text":1216,"level":247},{},{"id":1219,"data":3190,"type":361,"tunes":3206},{"content":3191,"stretched":43,"withHeadings":14},[3192,3193,3194,3195,3196,3197,3198,3199,3200,3201,3202,3203,3204,3205],[521,1223],[1225,1226],[1228,1229],[1231,1232],[1234,1235],[1237,1238],[1240,1241],[1243,1244],[1246,1247],[1249,1250],[1252,1253],[1255,1256],[1258,1259],[1261,1262],{},{"id":1265,"data":3208,"type":42,"tunes":3209},{"text":1267,"level":247},{},{"id":1270,"data":3211,"type":218,"tunes":3212},{"text":1272},{},{"id":1275,"data":3214,"type":218,"tunes":3215},{"text":1277},{},{"id":1280,"data":3217,"type":218,"tunes":3218},{"text":1282},{},{"id":1285,"data":3220,"type":218,"tunes":3221},{"text":1287},{},{"id":1290,"data":3223,"type":218,"tunes":3224},{"text":1292},{},{"id":1295,"data":3226,"type":42,"tunes":3227},{"text":1297,"level":247},{},{"id":1300,"data":3229,"type":218,"tunes":3230},{"text":1302},{},{"id":1305,"data":3232,"type":218,"tunes":3233},{"text":1307},{},{"id":1310,"data":3235,"type":218,"tunes":3236},{"text":1312},{},{"id":1315,"data":3238,"type":42,"tunes":3239},{"text":1317,"level":247},{},{"id":1320,"data":3241,"type":218,"tunes":3242},{"text":1322},{},{"id":1325,"data":3244,"type":218,"tunes":3245},{"text":1327},{},{"id":1330,"data":3247,"type":218,"tunes":3248},{"text":1332},{},{"id":1335,"data":3250,"type":42,"tunes":3251},{"text":1337,"level":247},{},{"id":1340,"data":3253,"type":1340,"tunes":3264},{"items":3254,"title":1379},[3255,3256,3257,3258,3259,3260,3261,3262,3263],{"id":1344,"answer":1345,"question":1346},{"id":1348,"answer":1349,"question":1350},{"id":1352,"answer":1353,"question":1354},{"id":1356,"answer":1357,"question":1358},{"id":1360,"answer":1361,"question":1362},{"id":1364,"answer":1365,"question":1366},{"id":1368,"answer":1369,"question":1370},{"id":1372,"answer":1373,"question":1374},{"id":1376,"answer":1377,"question":1378},{},{"id":1382,"data":3266,"type":42,"tunes":3267},{"text":1384,"level":247},{},{"id":1387,"data":3269,"type":1387,"tunes":3282},{"title":1389,"entries":3270},[3271,3272,3273,3274,3275,3276,3277,3278,3279,3280,3281],{"term":1392,"anchor":1393,"definition":1394},{"term":1396,"anchor":1397,"definition":1398},{"term":1400,"anchor":1401,"definition":1402},{"term":447,"anchor":1404,"definition":1405},{"term":1407,"anchor":1408,"definition":1409},{"term":1411,"anchor":1412,"definition":1413},{"term":1415,"anchor":1416,"definition":1417},{"term":1419,"anchor":1420,"definition":1421},{"term":1423,"anchor":1424,"definition":1425},{"term":1427,"anchor":1428,"definition":1429},{"term":1431,"anchor":1432,"definition":1433},{},{"id":1436,"data":3284,"type":42,"tunes":3285},{"text":1438,"level":247},{},{"id":1441,"data":3287,"type":218,"tunes":3288},{"text":1443},{},{"id":1446,"data":3290,"type":218,"tunes":3291},{"text":1448},{},{"id":1451,"data":3293,"type":218,"tunes":3294},{"text":1453},{},{"id":1456,"data":3296,"type":42,"tunes":3297},{"text":1458,"level":247},{},{"id":1461,"data":3299,"type":218,"tunes":3300},{"text":1463},{},{"id":1466,"data":3302,"type":1473,"tunes":3305},{"link":1468,"meta":3303},{"image":3304,"title":1471,"description":1472},{"url":347},{},{"id":1476,"data":3307,"type":1473,"tunes":3310},{"link":1478,"meta":3308},{"image":3309,"title":1481,"description":1482},{"url":347},{},{"id":1485,"data":3312,"type":1473,"tunes":3315},{"link":1487,"meta":3313},{"image":3314,"title":1490,"description":1491},{"url":347},{},{"id":1494,"data":3317,"type":1473,"tunes":3320},{"link":1496,"meta":3318},{"image":3319,"title":1499,"description":1500},{"url":347},{},{"id":1503,"data":3322,"type":1473,"tunes":3325},{"link":1505,"meta":3323},{"image":3324,"title":1508,"description":1509},{"url":347},{},{"id":1512,"data":3327,"type":1473,"tunes":3330},{"link":1514,"meta":3328},{"image":3329,"title":1517,"description":1518},{"url":347},{},{"id":1521,"data":3332,"type":1473,"tunes":3335},{"link":1523,"meta":3333},{"image":3334,"title":1526,"description":1527},{"url":347},{},{"id":1530,"data":3337,"type":1473,"tunes":3340},{"link":1532,"meta":3338},{"image":3339,"title":1535,"description":1536},{"url":347},{},"Post erfolgreich abgerufen",{"items":3343,"source":3428,"manualIds":3429,"manualMatchedIds":3430},[3344,3351,3358,3365,3372,3379,3386,3393,3400,3407,3414,3421],{"id":3345,"slug":3346,"title":3347,"excerpt":3348,"featuredImage":3349,"publishedAt":3350},"470","what-should-an-ai-agent-remember-forget-recompute-or-retrieve-again","Šta bi AI agent trebalo da zapamti, zaboravi, ponovo izračuna ili ponovo preuzme?","Dugotrajni agenti ne bi trebalo da pamte sve. Ovaj članak pruža praktičan model životnog ciklusa za odlučivanje o tome šta pripada trajnoj memoriji, šta bi trebalo ponovo preuzeti, šta je bezbednije ponovo izračunati i šta bi trebalo da istekne ili bude zamenjeno.","\u002Fuploads\u002F2026\u002F09\u002Fwhat-should-an-ai-agent-remember-forget-recompute-or-retrieve-again-1790351131087-iehz28.webp","2026-09-25T09:43:00.000Z",{"id":3352,"slug":3353,"title":3354,"excerpt":3355,"featuredImage":3356,"publishedAt":3357},"480","when-should-an-ai-stop-trusting-its-own-knowledge-the-retrieval-trigger","Kada bi AI trebalo da prestane da veruje sopstvenom znanju? — Okidač za pretragu","AI model ne zahteva pretragu za svako pitanje. Važan problem je znati kada njegovo interno znanje više nije dovoljno. Okidač za pretragu je praktična granica odlučivanja koja određuje kada AI sistem treba da prestane da se oslanja isključivo na znanje modela i pribavi spoljne dokaze pre odgovaranja.","\u002Fuploads\u002F2026\u002F09\u002Fwhen-should-an-ai-stop-trusting-its-own-knowledge-the-retrieval-trigger-1790574991244-f4rpyg.webp","2026-09-28T01:49:00.000Z",{"id":3359,"slug":3360,"title":3361,"excerpt":3362,"featuredImage":3363,"publishedAt":3364},"467","the-answer-validity-boundary-the-missing-layer-between-relevance-and-reliable-ai-answers","Granica valjanosti odgovora: Nedostajući sloj između relevantnosti i pouzdanih AI odgovora","Izvor može biti relevantan, autoritativan i ipak pogrešan za pitanje koje se postavlja. Sloj koji nedostaje je primenljivost: uslovi pod kojima odgovor važi i promene koje ga primoravaju na preispitivanje. Ovaj članak predstavlja Granicu važenja odgovora kao obrazac za dizajn izvora za ljude, AI pretragu i RAG sisteme.","\u002Fuploads\u002F2026\u002F09\u002Fthe-answer-validity-boundary-the-missing-layer-between-relevance-and-reliable-ai-answers-1790272901306-1g5jly.webp","2026-09-24T11:59:00.000Z",{"id":3366,"slug":3367,"title":3368,"excerpt":3369,"featuredImage":3370,"publishedAt":3371},"492","mcp-explained-what-it-connects-what-it-does-not-do-and-where-it-fits","MCP objašnjen: Šta povezuje, šta ne radi i gde se uklapa","Model Context Protocol povezuje AI aplikacije sa eksternim alatima, resursima i promptovima kroz standardnu granicu klijent-server. Saznajte šta MCP radi, šta ne radi i gde se uklapa u arhitekturu agenata.","\u002Fuploads\u002F2026\u002F10\u002Fmcp-explained-what-it-connects-what-it-does-not-do-and-where-it-fits-1791486640275-7ub1cq.webp","2026-10-08T15:09:00.000Z",{"id":3373,"slug":3374,"title":3375,"excerpt":3376,"featuredImage":3377,"publishedAt":3378},"486","source-of-truth-in-ai-systems-where-reliable-knowledge-actually-comes-from","Izvor istine u AI sistemima: Odakle pouzdano znanje zaista dolazi","Izvor istine definiše koji je izvor merodavan za određenu činjenicu ili stanje. Saznajte kako se razlikuje od RAG-a, porekla, memorije, konteksta, vektorskih baza podataka i sistema evidencije.","\u002Fuploads\u002F2026\u002F10\u002Fsource-of-truth-in-ai-systems-where-reliable-knowledge-actually-comes-from-1791479103235-6bq9em.webp","2026-10-08T13:02:00.000Z",{"id":3380,"slug":3381,"title":3382,"excerpt":3383,"featuredImage":3384,"publishedAt":3385},"487","vector-databases-embeddings-and-reranking-three-different-parts-of-retrieval","Vektorske baze podataka, ugrađivanja i ponovno rangiranje: Tri različita dela pretraživanja","Embedinzi predstavljaju značenje, vektorske baze podataka pronalaze kandidate, a rerangirači prečišćavaju rezultate. Saznajte kako se ova tri sloja pronalaženja razlikuju i kako rade zajedno u RAG-u.","\u002Fuploads\u002F2026\u002F10\u002Fvector-databases-embeddings-and-reranking-three-different-parts-of-retrieval-1791480129884-9dtasz.webp","2026-10-08T11:21:00.000Z",{"id":3387,"slug":3388,"title":3389,"excerpt":3390,"featuredImage":3391,"publishedAt":3392},"495","sovereign-ai-control-of-models-data-infrastructure-and-dependencies","Suverena AI: Kontrola modela, podataka, infrastrukture i zavisnosti","Suverena AI se odnosi na efektivnu kontrolu nad modelima, podacima, infrastrukturom, softverom, operacijama i strateškim zavisnostima — a ne samo na to gde je AI model hostovan.","\u002Fuploads\u002F2026\u002F10\u002Fsovereign-ai-control-of-models-data-infrastructure-and-dependencies-1791488833132-niy85x.webp","2026-10-08T15:45:00.000Z",{"id":3394,"slug":3395,"title":3396,"excerpt":3397,"featuredImage":3398,"publishedAt":3399},"466","the-gpu-is-not-the-product-future-proof-private-ai-architecture","GPU nije proizvod: Privatna AI arhitektura spremna za budućnost","Privatna AI infrastruktura ne bi trebalo da bude projektovana oko jednog GPU-a ili jednog modela. Otporniji pristup kombinuje brze GPU-ove za inferenciju, memorijski bogate AI sisteme, čvorove za fizički AI i opcione vodeće modele u oblaku iza sloja za rutiranje koji prepoznaje mogućnosti.","\u002Fuploads\u002F2026\u002F09\u002Fthe-gpu-is-not-the-product-future-proof-private-ai-architecture-1790140878812-8hsl39.webp","2026-09-23T01:19:00.000Z",{"id":3401,"slug":3402,"title":3403,"excerpt":3404,"featuredImage":3405,"publishedAt":3406},"483","what-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs","Šta je AI rešenje arhitekta? Granice sistema, odgovornosti i kompromisi","AI Solution Architect pretvara poslovne zahteve u AI sistem spreman za produkciju, obuhvatajući podatke, modele, alate, bezbednost, izvršno okruženje, evaluaciju i operacije.","\u002Fuploads\u002F2026\u002F10\u002Fwhat-is-an-ai-solution-architect-system-boundaries-responsibilities-and-trade-offs-1791476643267-1st5xz.webp","2026-10-08T12:23:00.000Z",{"id":3408,"slug":3409,"title":3410,"excerpt":3411,"featuredImage":3412,"publishedAt":3413},"490","rbac-vs-tenant-isolation-two-different-security-boundaries","RBAC naspram izolacije zakupaca: dve različite bezbednosne granice","RBAC kontroliše šta korisnik sme da radi; izolacija zakupaca kontroliše kojim resursima tog zakupca ta radnja može da pristupi. Saznajte zašto bezbednost višekorisničkog SaaS-a zahteva obe granice.","\u002Fuploads\u002F2026\u002F10\u002Frbac-vs-tenant-isolation-two-different-security-boundaries-1791485111528-qqtzby.webp","2026-10-08T14:43:00.000Z",{"id":3415,"slug":3416,"title":3417,"excerpt":3418,"featuredImage":3419,"publishedAt":3420},"468","ai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context","Memorija AI agenta nije RAG: Kako razdvojiti memoriju, pronalaženje, stanje i kontekst","Memorija agenta, RAG, stanje i kontekst često se koriste kao da su međusobno zamenjivi. Oni to nisu. Ovaj praktični arhitektonski model razdvaja ova četiri sloja, pokazuje gde svaki pripada i objašnjava šta se kvari kada ih sistemi stope u jedno.","\u002Fuploads\u002F2026\u002F09\u002Fai-agent-memory-is-not-rag-how-to-separate-memory-retrieval-state-and-context-1790350560308-np0xy6.webp","2026-09-25T11:34:00.000Z",{"id":3422,"slug":3423,"title":3424,"excerpt":3425,"featuredImage":3426,"publishedAt":3427},"484","what-is-an-ai-platform-architect-models-data-runtime-security-and-operations","Šta je arhitekta AI platforme? Modeli, podaci, izvršno okruženje, bezbednost i operacije","Arhitekta AI platforme projektuje višekratno upotrebljive AI temelje kroz modele, provajdere, pretragu, agente, identitet, bezbednost, evaluaciju, opservabilnost i operacije.","\u002Fuploads\u002F2026\u002F10\u002Fwhat-is-an-ai-platform-architect-models-data-runtime-security-and-operations-1791477229171-ou3zcc.webp","2026-10-08T12:32:00.000Z","fallback",[],[]]