{"id":4628,"date":"2026-10-23T03:59:00","date_gmt":"2026-10-23T03:59:00","guid":{"rendered":"https:\/\/11000.dev\/index.php\/2026\/10\/23\/microsoft-predstavio-thinkingbox-ai-agenti-se-sada-ocenjuju-po-tome-sta-stvarno-urade-u-bazi\/"},"modified":"2026-10-23T03:59:00","modified_gmt":"2026-10-23T03:59:00","slug":"microsoft-predstavio-thinkingbox-ai-agenti-se-sada-ocenjuju-po-tome-sta-stvarno-urade-u-bazi","status":"publish","type":"post","link":"https:\/\/11000.dev\/index.php\/2026\/10\/23\/microsoft-predstavio-thinkingbox-ai-agenti-se-sada-ocenjuju-po-tome-sta-stvarno-urade-u-bazi\/","title":{"rendered":"Microsoft predstavio ThinkingBox: AI agenti se sada ocenjuju po tome \u0161ta stvarno urade u bazi"},"content":{"rendered":"<p><figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/11000.dev\/wp-content\/uploads\/2026\/10\/oS8I8qx7iBS8EcNIXR3q9.png\" alt=\"Microsoft predstavio ThinkingBox: AI agenti se sada ocenjuju po tome \u0161ta stvarno urade u bazi\" \/><\/figure>\n<\/p>\n<p>Microsoft je na Hugging Face platformi objavio novi sistem pod nazivom <strong>ThinkingBox<\/strong>, koji menja na\u010din na koji se treniraju i ocenjuju AI agenti. Umesto da se uspeh meri samo po tome da li je model &#8220;rekao&#8221; da je zadatak zavr\u0161en, ThinkingBox uvodi <strong>RL (reinforcement learning) nagrade zasnovane na stvarnom stanju baze podataka<\/strong> \u2014 drugim re\u010dima, agent dobija poene tek kada se podaci u bazi zaista promene onako kako je tra\u017eeno.<\/p>\n<p>Ovo zvu\u010di jednostavno, ali re\u0161ava jedan od najve\u0107ih problema dana\u0161njih agenata: sklonost ka tome da tvrde da su posao zavr\u0161ili, dok u pozadini ni\u0161ta nije ura\u0111eno. Uz objavu je prilo\u017een i rad pod naslovom <em>&#8220;One Success Isn&#8217;t Reliability&#8221;<\/em>, koji eksplicitno upozorava da jedan uspe\u0161an poku\u0161aj ne zna\u010di pouzdanost \u2014 agent mora da bude konzistentan kroz mnoge iteracije u realnim, stateful poslovnim tokovima.<\/p>\n<h2>\u0160ta ta\u010dno donosi ThinkingBox<\/h2>\n<p>U srcu sistema nalazi se <strong>sandbox okru\u017eenje<\/strong> u kome agent izvr\u0161ava zadatke nad bazom podataka, a zatim se rezultat automatski proverava. Ako je stanje baze ispravno \u2014 nagrada sledi. Ako nije, sledi korekcija. Time se proces treniranja pomera sa &#8220;jezi\u010dkog ube\u0111ivanja&#8221; na <strong>merljive, proverljive ishode<\/strong>.<\/p>\n<p>Microsoft je uz to objavio i <strong>ThinkingBox-Bench<\/strong>, skup zadataka i metrika za testiranje agenata u scenarijima koji li\u010de na stvarne poslovne workflow-e \u2014 tamo gde gre\u0161ka u bazi mo\u017ee da zna\u010di pravi problem, a ne samo lo\u0161 odgovor u chatu. Benchmark je dostupan na Hugging Face-u i ve\u0107 bele\u017ei solidnu pa\u017enju zajednice.<\/p>\n<h2>Za\u0161to je ovo va\u017eno za srpski tech ekosistem<\/h2>\n<p>Za doma\u0107e timove koji razvijaju AI agente za automatizaciju poslovnih procesa, vest je vi\u0161e nego relevantna. Pouzdanost je ta koja odvaja demo projekat od proizvoda koji mo\u017ee da se pusti u produkciju. Pristup &#8220;proveri stanje, ne izjavu&#8221; mogao bi da postane standard u razvoju agenata koji rade sa <strong>bazama podataka, ERP sistemima i internim alatima<\/strong>.<\/p>\n<p>Vi\u0161e detalja o metodu i samom benchmark-u mo\u017eete prona\u0107i u <a href=\"https:\/\/huggingface.co\/blog\/microsoft\/thinkingbox-training\" target=\"_blank\" rel=\"noopener\">originalnom Microsoft blog postu na Hugging Face-u<\/a>.<\/p>\n<p>Ovo je jo\u0161 jedan signal da se trka AI agenata polako seli sa pitanja &#8220;\u0161ta model mo\u017ee da ka\u017ee&#8221; na pitanje &#8220;\u0161ta model mo\u017ee da uradi \u2014 i doka\u017ee&#8221;.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Microsoft je na Hugging Face platformi objavio novi sistem pod nazivom ThinkingBox, koji menja na\u010din na koji se treniraju i ocenjuju AI agenti. Umesto da se uspeh meri samo po tome da li je model &#8220;rekao&#8221; da je zadatak zavr\u0161en, ThinkingBox uvodi RL (reinforcement learning) nagrade zasnovane na stvarnom stanju baze podataka \u2014 drugim re\u010dima, [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":4627,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"content-type":"","footnotes":""},"categories":[103,104],"tags":[],"class_list":["post-4628","post","type-post","status-publish","format-standard","has-post-thumbnail","category-ai","category-tehnologija"],"_links":{"self":[{"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/posts\/4628","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/comments?post=4628"}],"version-history":[{"count":0,"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/posts\/4628\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/media\/4627"}],"wp:attachment":[{"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/media?parent=4628"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/categories?post=4628"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/11000.dev\/index.php\/wp-json\/wp\/v2\/tags?post=4628"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}