Yawancin zagayen agent suna gudanar da tool-call da zarar model ya gabatar da shi, kuma sakamakon JevAct na nuna cewa wurin dubawa mai arha, bisa yiwuwa, da aka sa kafin aiwatarwa na iya ƙara amincin agents masu amfani da kayan aiki, ko da yake ribar ta bambanta ƙwarai daga benchmark zuwa benchmark.
Taƙaitawa ga Shugabanni
- LLM agents suna ƙara ɗaukar matakai masu tasiri ta kayan aiki na waje, amma ana yawan aiwatar da waɗannan ayyuka ba tare da tantancewa mai zaman kanta ta bin ƙa'idoji ko kammala aiki ba.
- JevAct yana fassara ƙa'idar fanni a waje da aiki zuwa tambayoyi, dokoki, binciken yanayi da ake iya aiwatarwa, kuma inda ake buƙata, jerin dubawa na mika aiki.
- Yayin aiki, code yana duba gaskiya a bayanan kayan aiki yayin da Kev, decision model irin na Jev da aka buɗe tushensa, ke amsa tambayoyin da suka dogara kan tattaunawa.
- Idan aka kwatanta da ReAct harness a τ²-bench (Airline, Telecom) da BFCL, riba ta daidaita a Telecom da nasarar gudu uku ta BFCL, kuma ta fi gauraya a Airline.
- Tsarin marubutan kansu na taka-tsantsan ne: sakamakon yana nuna cewa decision models masu tsari tare da tantancewa da ake iya aiwatarwa na iya ƙara aminci.
Matsalar: aiwatarwa na bin gabatarwa #
Agent da ke yin ajiyar tikitin jirgi, soke oda ko canza saitin asusu yana yin abin da ba za a iya warwarewa a ɓoye ba. Abin da ba shi da daɗi shi ne ƙarancin abin da ke tsaye tsakanin shawarar model da tasirinta. Yawancin zagayen agent suna aiwatar da tool-call da aka gabatar da zarar model ya samar da shi. Babu wanda ke duba, a zaman kansa, ko aikin ya bi ƙa'idar fannin, ko aikin da agent ya ɗauka ya kammala ya kammala da gaske.
Wannan giɓi shi ne mafarin "Think Before You Act: Decision Gate for LLM Agents", wani preprint na arXiv na 2026 daga Iftach Shoham, Yuval David da abokan aikinsu a IBM, wanda ya gabatar da JevAct. Wannan ba sabon lura ba ne, amma ya dace a faɗe shi a fili domin yawancin kayan aikin agent suna mai da hankali kan inganta mai gabatarwa. ReAct agents suna haɗa tunani da amfani da kayan aiki, kuma wannan salo ya zama mizani na gama-gari. Hanyoyin tunani-baya irin su Self-Refine da Reflexion suna ƙara suka da ra'ayi don inganta gabatarwa ko gwaje-gwaje na gaba, ba tare da sabunta nauyin model ba. Duk amfaninsu, waɗannan hanyoyin ba sa, da kansu, tantance wani takamaiman aiki kan ƙa'idar fanni daidai kafin aiwatarwa. Binciken da JevAct ke damuwa da shi yana wani wuri daban a zagayen: bayan gabatarwa ta wanzu kuma kafin ta tabbata.
Menene JevAct #
JevAct harness ne da ke duba zaɓaɓɓun gabatarwa kafin aiwatarwa. An bayyana shi a matsayin tsarin runtime verification da ke tare zaɓaɓɓun ayyukan agent kuma yana ba da damar gyara bisa ra'ayi. Yana aiki ba tare da dogaro kan model na agent da ke ƙasa ba kuma baya buƙatar canza sigogin model ko kayan aikinsa. A tsarin da aka kimanta, LLM na agent, babban prompt ɗinsa da kayan aikinsa ba su canza ba.
A tsarin gine-gine, JevAct yana ƙara wa harness na ReAct agent ta saka matakin dubawa na zaɓi tsakanin gabatar da saƙo da aiwatarwa. Kalmar "na zaɓi" na da muhimmanci. Tsarin ba ya duba kowane saƙo; yana nufin takamaiman nau'ikan aiki. An ayyana biyu a takardar. Write call yana canza muhalli, misali ta ajiya, soke wani abu ko canza saiti. Hand-off yana kawo ƙarshen ƙoƙarin agent, ko ta mika mai amfani ga mutum ko ta gaya wa mai amfani cewa babu abin da za a ƙara yi. A writes ne keta ƙa'ida ke yin ɓarna; a hand-offs ne agent zai iya haƙura da wuri. Manufar tsarin da aka bayyana, ta gano ayyukan da ka iya zama marasa inganci da dakatar da aiki kafin lokaci, ta yi daidai da waɗannan rukunoni biyu.
Idan dubawa ta kasa, ba a toshe gabatarwar kawai. Binciken na iya jinkirta aiwatarwa kuma ya nemi gabatarwa da aka gyara daga LLM tare da ra'ayi na ɗan lokaci, ƙarƙashin ƙayyadadden kasafin gyara. Gabatarwar da ta wuce tana ci gaba. Wadda aka yi wa alama tana komawa ga agent tare da ra'ayi na musamman da damar gyara mai iyaka. Kuma rabon aiki yana nan a fili: agent ne har yanzu ke gabatar da ayyuka, kuma Kev ba ya zaɓar su.
Iri biyu na shaida, ƙofa ɗaya #
Ƙirar ta tsaya kan rarrabuwa tsakanin abin da za a iya tantancewa da tabbas da abin da ba za a iya ba. JevAct yana haɗa tantancewa mai tabbas ta gaskiyar yanayin kayan aiki da kimantawa bisa yiwuwa ta yanayin tattaunawa. Yayin aiki, code yana duba gaskiyar da ke cikin bayanan kayan aiki, kuma Kev yana amsa tambayoyin da suka dogara kan tattaunawa.
Wannan rarrabuwa na da ma'ana. Wasu sharuɗɗan ƙa'ida gaskiya ne da ke zaune a bayanan kayan aiki, kuma manhaja za ta iya karanta su ba tare da ra'ayin kowane model ba. Wasu sharuɗɗan suna cikin abin da mai amfani da agent suka faɗa wa juna, wanda babu binciken database da zai warware shi. Bai wa code nau'in farko da model nau'i na biyu yana nufin ana tambayar model ne kawai game da abin da code ba zai iya amsawa ba.
Rabin da ke waje da aiki shi ne ke sa a iya bayyana wannan a aikace. Compiler yana mayar da ƙa'idar fanni zuwa tambayoyi, dokoki, binciken da ake iya aiwatarwa kan yanayin kayan aiki, kuma inda ake buƙata, jerin dubawa na hand-off. A wata ma'ana, JevAct yana fassara ƙa'idojin fanni zuwa binciken yanayi da ake iya aiwatarwa da tambayoyi masu tsari da decision model mai prefill kawai ke kimantawa. Marubutan sun ce sun aza harsashin tsarin aiki da ke fassara ƙa'idojin fanni zuwa tambayoyi, dokokin yanke shawara, binciken yanayi da ake iya aiwatarwa da jerin dubawa na hand-off don ƙofar decision model. Wannan hanya ce mai sassauci ta faɗa, kuma zan karɓe ta yadda take: iƙirarin game da harsashin tsarin aiki ne, ba samfurin da ya kammala ba.
Me ya sa decision model maimakon alƙalin LLM #
Madadin da ke bayyane shi ne alƙalin LLM ya karanta gabatarwar ya rubuta hukunci. Ƙin yarda na takardar na aiki ne. Bayanin da irin wannan alƙali ya rubuta a sake dole a fassara shi kafin ƙofa ta iya aiki da shi, kuma yanke wannan shawara da frontier model yana da tsada. Ƙofa tana buƙatar shawara, ba maƙala ba.
An gina decision models don haka. Jev ya gabatar da decision models da ke mayar da zaɓuɓɓuka masu nau'i, maki da yiwuwa maimakon rubutun da aka samar. Suna mayar da zaɓuɓɓuka da yiwuwa ga tambayoyi ƙayyadaddu, wanda ke sa binciken bisa iyaka ya zama mai sauƙi. Model ɗin da JevAct ke amfani da shi mai prefill kawai ne kuma yana samar da hukunci bisa yiwuwa ba tare da dogaro kan samar da rubutu na LLM ba. Kev decision model ne irin na Jev mai buɗaɗɗen tushe, wanda aka sa a matakin tantancewa daban kafin zaɓaɓɓun ayyuka su tabbata.
Akwai wata matsala, kuma marubutan sun faɗe ta da kansu. Ba a da tabbacin cewa yiwuwar decision model an daidaita ta (calibrated) ko daidai take, don haka dole a gwada amfaninta a zagayen agent maimakon a ɗauka. Wannan jimla ce ke tsara dukan ɓangaren gwaji. Babban tambayar ita ce ko decision model zai iya aiki a matsayin mai tantancewa mai iyakar yiwuwa tare da binciken yanayin kayan aiki da ake iya aiwatarwa. Tambaya ce da gwaji zai amsa, kuma a gwajin ne cikakkun bayanai masu ban sha'awa suke.
Ayyukan da suka shafi decision models suna ba da hoto mai gauraye na dalilin da ya sa wannan gwaji ke da muhimmanci. Li da abokan aikinsa sun ruwaito sakamako mai gasa a kan fifiko da gaskiya mai tushe a shaida, amma aiki mai rauni idan ana buƙatar fitar da sakamako ta matakai. Rao da Callison-Burch sun ruwaito sakamakon rubric mai gauraye da kurakurai da suka zo ɗaya da na alƙalan LLM. Decision model da ke da ƙarfi a wasu ayyukan hukunci kuma mai rauni a wasu shi ne ainihin irin ɓangaren da za ka so ka haɗa da binciken mai tabbas maimakon ka amince da shi shi kaɗai.
Yadda aka gudanar da kwatancen #
An kwatanta JevAct da ReAct harness ta amfani da LLM ɗaya na agent a ɓangarorin biyu. Ayyukan sun haɗa da Airline da Telecom a τ²-bench da ayyukan function-calling masu juyi da yawa a BFCL. Riƙe LLM na agent iri ɗaya a ɓangarorin biyu shi ne iko daidai ga tambayar da ake yi: duk wani bambanci yana fitowa daga harness, ba daga mai gabatarwa mafi ƙarfi ko mafi rauni ba.
Takardar ta ruwaito sakamakonta ta amfani da alkaluman pass^3, daga Table 2, tare da nasara aƙalla sau ɗaya. Waɗannan ma'auni biyu suna ja zuwa alkibla daban-daban, kuma bambancinsu ya zama mai ba da bayani. Nasara a dukan gudu uku tana ba da lada ga agent da ke yin daidai kowane lokaci. Nasara aƙalla sau ɗaya tana ba da lada ga agent da ya yi daidai a kowane ƙoƙari. Mai tantancewa da ke kama ayyuka marasa kyau ya kamata ya taimaki na farko fiye da na biyu da tabbas, kuma sakamakon ya dace da wannan karatu, ko da yake takardar ta ce ne kawai salon BFCL ya dace da aiki mafi maimaituwa a wasu ayyuka, don haka zan nuna sauran a matsayin fassara, ba sakamako ba.
Abin da alkaluman ke nunawa #
A Telecom, dukan agents takwas sun inganta duka nasara aƙalla sau ɗaya da nasara a dukan gudu uku. Wannan shi ne sakamako mafi tsabta a takardar: babu keɓe a cikin agents da aka gwada, a kowane ma'auni.
A BFCL, dukan agents tara sun inganta a nasara a gudu uku, amma sakamakon aƙalla sau ɗaya ya gauraya. A karanta wannan a hankali. A BFCL, ma'auni mafi tsauri ya tashi ga kowane agent da aka gwada, da maki 0.1 zuwa 9.6; ma'auni mai sassauci ya inganta ga agents huɗu kuma ya ragu ga biyar. Idan ka damu da agent ya yi daidai kowane lokaci, wannan alkibla ce mai kyau. Idan ka damu da ko sake gwadawa zai iya yin nasara daga ƙarshe, ba a tabbatar da ribar a nan ba.
Airline shi ne wuri mai rauni. Tasirin Airline ya fi ƙanƙanta kuma ya bambanta daga agent zuwa agent. Taƙaitaccen sakamakon ya faɗa a matsayin riba mai daidaito musamman a Telecom da BFCL, yayin da aiki a Airline ya fi gauraya. Wanda ya kafa kamfani da ke karanta wannan bai kamata ya ƙara shi zuwa "yana inganta komai" ba. Marubutan sun ruwaito ingantuwar amincin agent a models da benchmarks da yawa, amma a Airline ribar ta fi ƙanƙanta kuma ta gauraya, kuma sun ce sakamakonsu bai tabbatar da riba iri ɗaya ba. Takardar ta yi gargaɗi cewa da ayyukan Airline 50, aiki ɗaya yana motsa maki da maki 2 na kashi, don haka a karanta ƙananan bambance-bambance da taka-tsantsan; bayan haka, ba zan yi hasashen dalilin da ya sa Airline ke nuna hali daban ba.
Marubutan kansu sun yi taka-tsantsan a karatun gaba ɗaya: sakamakon yana nuna cewa haɗa decision models masu tsari da tantancewa da ake iya aiwatarwa na iya ƙara amincin LLM agents masu amfani da kayan aiki. "Yana nuna" shi ne fi'ili daidai. Sakamako ne mai alƙawari, ba hujja ba.
Abin da kwatancen ke ware da wanda ba ya ware #
Iyaka ɗaya ta cancanci nauyi fiye da wanda ake ba ta. Domin JevAct yana ƙara binciken da aka fassara da damar gyara, kwatancen suna kimanta cikakken harness, ba alƙalin shi kaɗai ba. Wannan gargaɗi ne mai muhimmanci ga duk wanda ke fatan ƙarasa cewa decision model ne ke ɗaukar babban nauyin. Ribar na iya fitowa daga yiwuwar Kev, daga binciken yanayi da ake iya aiwatarwa, daga ƙarin damar gyara, ko daga wani haɗi. Manyan kwatancen ba sa raba su.
Zan ce wannan ba aibi ba ne sosai, illa bayanin gaskiya na abin da takardar tsari za ta iya nunawa. Tambayar da marubutan suka yi, ko decision model zai iya aiki a matsayin mai tantancewa mai iyakar yiwuwa tare da binciken yanayin kayan aiki da ake iya aiwatarwa, an amsa ta a matakin biyun tare, tunda "tare da" yana cikin tambayar kanta. Amma mai karatu da ke yanke shawarar inda zai zuba ƙoƙari ya kamata ya riƙe wannan a rai. Idan kana gina wani abu makamancin haka, ba za ka iya ƙarasa daga wannan shaida cewa decision model shi kaɗai zai ɗauki ingantuwar ba, ko kuma binciken code shi kaɗai.
Hakan kuma na nufin gargaɗin daidaitawa yana nan. Ba a da tabbacin cewa yiwuwar an daidaita ta ko daidai take, kuma sakamakon matakin harness bai kawar da wannan rashin tabbas ba. Yana nuna cewa tsarin gaba ɗaya ya taimaka a benchmarks da aka gwada.
Inda wannan yake cikin sauran ƙirar kariya #
JevAct ya shiga cikin tarin hanyoyi masu yawa na kare ayyukan agent, kuma bambance-bambancen suna koyarwa. GuardAgent yana amfani da agent daban don samarwa da aiwatar da code na tantancewa daga buƙatun kariya da mai amfani ya ayyana. Zwerdling da abokan aikinsa suna fassara ƙa'idojin kamfani zuwa kariya da ake iya aiwatarwa da ke duba bin ƙa'ida kafin ayyukan kayan aiki. AgentSpec yana wakiltar ƙuntatawa ta lokacin aiki a matsayin abubuwan jawowa, predicates da ayyukan tilastawa. ShieldAgent yana fitar da dokoki da ake iya tantancewa daga takardun ƙa'ida.
Wasu suna dogaro kan models don tantancewar kanta. ToolSafe yana amfani da guardrail da aka koyar don kimanta tool-calls da aka gabatar da mayar da ra'ayi, yayin da PolicyGuard da PolicyGuide ke amfani da masu tantancewa bisa model don bin ƙa'ida da jagorar tsarin aiki. CaMeL ya bi wata hanya daban, yana tilasta tsaro ta raba kwararar sarrafawa amintacciya da bayanai marasa amana.
A kan wannan yanayi, zaɓin da ya bambanta JevAct shi ne haɗinsa na musamman. Kamar layin aikin fassara zuwa kariya da ake iya aiwatarwa, yana mayar da ƙa'ida zuwa binciken yanayin kayan aiki. Kamar masu tantancewa bisa model, yana amfani da model don hukuncin da code ba zai iya yi ba. Abin da yake ƙarawa shi ne decision model da ke mayar da yiwuwar zaɓi ƙayyadadde, da ake amfani da shi kai tsaye kafin zaɓaɓɓun ayyuka su aiwatu, tare da ra'ayi don gyara. Yana barin gabatarwa ga LLM na ReAct kuma yana sa Kev a matakin tantancewa daban kafin zaɓaɓɓun ayyuka su tabbata.
Me ya sa wannan ƙira ta cancanci a bi ta #
A ganina ra'ayi mafi amfani a nan ba Kev kansa ba ne, illa tsarin rarrabuwar. A tambayi code abin da code zai iya amsawa. A tambayi model abin da ya dogara kan tattaunawa kawai. A raba mai gabatarwa da mai tantancewa, ta yadda model da ke duba aiki ba shi ne ya ba da shawarar ba. Kuma a sa sakamakon mai tantancewa ya zama yiwuwa a kan iyaka, ta yadda ƙofar ta zama kwatanci maimakon rubutu da wani zai fassara.
Wannan ƙira kuma tana sa wuraren rauni su bayyana. Idan dubawa ta yi kuskure, za ka iya tambaya ko binciken yanayi ne ko tambayar tattaunawa. Idan iyakokin ba daidai ba ne, lambobi ne da za ka iya bincika. Gargaɗin takardar kanta cewa yiwuwa na buƙatar gwaji a cikin zagayen shi ne hankali daidai, kuma matsayi ne mafi kyau fiye da ɗaukar cewa mai tantancewa yana aiki domin yana da ma'ana a ji.
Sakamakon da ba daidai ba, idan ma akwai komai, yana ƙarfafa amincewa da gaskiyar rahoton. Tsarin da ke iƙirarin riba iri ɗaya a ko'ina zai cancanci ƙarin shakka fiye da wanda ke nuna takwas cikin takwas a Telecom, tara cikin tara a nasarar gudu uku ta BFCL, sakamakon aƙalla sau ɗaya mai gauraye a BFCL, da tasiri mai bambanta a Airline. Salon yana bai wa masu aiki abin da za su tsara a kai: shaida mafi ƙarfi ita ce ta aminci a gudu masu maimaituwa, kuma mafi rauni tana a fannin da tasiri ke bambanta daga agent zuwa agent.
Abin da za a yi da wannan #
Ga masu karatu na fasaha, darasin na tsarin gine-gine ne. LLM na agent, babban prompt ɗinsa da kayan aikinsa ba sa canzawa a JevAct, don haka matakin tantancewa abu ne da kake ƙarawa kewaye da agent, ba abin da kake koyar da shi a ciki ba. Ga waɗanda suka kafa kamfanoni, karatun aiki ya fi ƙunci fiye da kanun labarin. Takardar ta ruwaito ingantaccen aminci a models da benchmarks da yawa, tare da Airline mafi gauraye, kuma kwatancenta suna kimanta cikakken harness maimakon alƙalin shi kaɗai. Manyan kwatancen ba su tabbatar da wane ɓangare ne ke samun ingantuwar ba, ko da yake ablation na Airline a takardar yana nuna cewa tantancewa ta manhaja ce ke da yawancinta a wannan yanayi.
Tambayar da ke buɗe ita ce babbar tambayar takardar kanta: ko decision model zai iya aiki a matsayin mai tantancewa mai iyakar yiwuwa tare da binciken yanayin kayan aiki da ake iya aiwatarwa, alhali, kamar yadda marubutan suka lura, ba a da tabbacin yiwuwarsa an daidaita ta ko daidai take. Sakamakon Airline, inda tasiri ya fi ƙanƙanta kuma ya bambanta daga agent zuwa agent, shi ne wurin da za a fara dubawa. Ko wannan bambanci ya samo asali daga binciken, iyakokin ko ƙa'idojin da ake fassarawa, abu ne da sakamakon da aka ruwaito ya bar a buɗe, kuma shi ne gwajin da na fi so in gani na gaba.
Tambayoyin da Ake Yawan Yi #
Menene JevAct? JevAct tsarin runtime verification ne daga masu bincike na IBM da ke tare zaɓaɓɓun ayyukan agent, wato write calls da hand-offs, kafin aiwatarwa. Gabatarwar da aka yi wa alama tana komawa ga agent tare da ra'ayi na musamman da damar gyara mai iyaka. LLM na agent, babban prompt ɗinsa da kayan aikinsa ba sa canzawa.
Me Kev ke yi a JevAct? Kev decision model ne irin na Jev mai buɗaɗɗen tushe da ke amsa tambayoyi ƙayyadaddu da yiwuwa maimakon rubutun da aka samar. A JevAct yana amsa tambayoyin da suka dogara kan tattaunawa, yayin da code ke duba gaskiyar da ke cikin bayanan kayan aiki. Kev ba ya zaɓar ayyuka; agent ne har yanzu ke gabatar da su.
Yaya aka kimanta JevAct? Takardar ta kwatanta JevAct da ReAct harness ta amfani da LLM ɗaya na agent a ɓangarorin biyu. Ayyukan sun haɗa da Airline da Telecom a τ²-bench da ayyukan function-calling masu juyi da yawa a BFCL, tare da ruwaito nasara a matsayin aƙalla sau ɗaya a cikin gudu uku da kuma nasara a dukan gudu uku.
Shin JevAct ya inganta kowane agent a kowane benchmark? A'a. A Telecom, dukan agents takwas sun inganta a ma'auni biyu, kuma a BFCL dukan tara sun inganta a nasara a gudu uku yayin da sakamakon aƙalla sau ɗaya ya gauraya. Tasirin Airline ya fi ƙanƙanta kuma ya bambanta daga agent zuwa agent, kuma marubutan sun ce sakamakon bai tabbatar da riba iri ɗaya ba.
Shin takardar ta nuna cewa decision model shi kaɗai ne ke kawo ribar? A'a. Domin JevAct yana ƙara binciken da aka fassara da damar gyara, manyan kwatancen suna kimanta cikakken harness, ba alƙalin shi kaɗai ba. Marubutan sun kuma lura cewa ba a da tabbacin cewa yiwuwar decision model an daidaita ta ko daidai take.
Manazarta #
- Shoham, I., David, Y., Akrabi, O., Bilgory, E., Levy, I., Marreed, S., Zeltyn, S. and Shlomov, S., (2026). Think Before You Act: Decision Gate for LLM Agents ⧉. arXiv.
- Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K. and Cao, Y., (2022). ReAct: Synergizing Reasoning and Acting in Language Models ⧉. arXiv.
- Barres, V., Dong, H., Ray, S., Si, X. and Narasimhan, K., (2025). τ²-bench: Evaluating Conversational Agents in a Dual-Control Environment ⧉. arXiv.
- TypeSafe AI, (2026). Gabatar da System One Models da Jev ⧉. TypeSafe AI.
- Palmer, J., (2026). KEV: Decision Models masu Prefill Kawai ⧉. GitHub.
- Xiang, Z. et al., (2024). GuardAgent: Safeguard LLM Agents by a Guard Agent via Knowledge-Enabled Reasoning ⧉. arXiv.
- Chen, Z., Kang, M. and Li, B., (2025). ShieldAgent: Shielding Agents via Verifiable Safety Policy Reasoning ⧉. arXiv.
- Li, Y., Miao, Y., Krishnan, R. and Padman, R., (2026). Jev-as-a-Judge: Accept When Confident, Escalate When Unsure ⧉. arXiv.
- Rao, D. and Callison-Burch, C., (2026). Jev vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places ⧉. arXiv.
Bita ta ƙarshe .
Sake buga wannan labarin
Kwafa tsarin Medium
# Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau > Originally published at [https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/](https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/) JevAct na duba zaɓaɓɓun ayyukan LLM agents kafin su gudana. Kan ReAct mai LLM ɗaya, riba ta daidaita a Telecom da BFCL amma ta gauraya a Airline. Read the full article on sebastienrousseau.com: https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/
Kwafa tsarin Mastodon
Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau JevAct na duba zaɓaɓɓun ayyukan LLM agents kafin su gudana. Kan ReAct mai LLM ɗaya, riba ta daidaita a Telecom da BFCL amma ta gauraya a Airline. https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/
Kwafa an tsara don LinkedIn
Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau JevAct na duba zaɓaɓɓun ayyukan LLM agents kafin su gudana. Kan ReAct mai LLM ɗaya, riba ta daidaita a Telecom da BFCL amma ta gauraya a Airline. Ga abubuwan da ya kamata a lura da su na dabarun: - Matsalar: aiwatarwa na bin gabatarwa. Agent da ke yin ajiyar tikitin jirgi, soke oda ko canza saitin asusu yana yin abin da ba za a iya warwarewa a ɓoye ba. - Menene JevAct. JevAct harness ne da ke duba zaɓaɓɓun gabatarwa kafin aiwatarwa. - Iri biyu na shaida, ƙofa ɗaya. Ƙirar ta tsaya kan rarrabuwa tsakanin abin da za a iya tantancewa da tabbas da abin da ba za a iya ba. - Me ya sa decision model maimakon alƙalin LLM. Madadin da ke bayyane shi ne alƙalin LLM ya karanta gabatarwar ya rubuta hukunci. Menene hanyar ƙungiyar ku wajen magance ƙalubalen da aka kawo a wannan rubuce-rubucen? → https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/ #LlmAgents #TantanceAyyukanAgent #RuntimeVerification #DecisionModels #AgentsMasuAmfaniDaKayanAiki Sebastien Rousseau | CC-BY-4.0
Buga wannan labari
Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau
JevAct na duba zaɓaɓɓun ayyukan LLM agents kafin su gudana. Kan ReAct mai LLM ɗaya, riba ta daidaita a Telecom da BFCL amma ta gauraya a Airline.
BibTeX
@online{rousseau2026tantance,
author = {Rousseau, Sebastien},
title = {{Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau}},
year = {2026},
url = {https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/},
urldate = {2026}
}RIS
TY - GEN AU - Rousseau, Sebastien TI - Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau PY - 2026 UR - https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/ ER -
Vancouver
Rousseau S. Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau. sebastienrousseau.com. 2026 Nov 18. Available from: https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/
Chicago
Rousseau, Sebastien. "Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau." sebastienrousseau.com. November 18, 2026. https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/.
APA
Rousseau, S. (2026, November 18). Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau. sebastienrousseau.com. https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/
Sake buga wannan labari
Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau
JevAct na duba zaɓaɓɓun ayyukan LLM agents kafin su gudana. Kan ReAct mai LLM ɗaya, riba ta daidaita a Telecom da BFCL amma ta gauraya a Airline.
An lasisin wannan labari a karkashin Creative Commons Attribution 4.0 International. Sake bugawa na bukatar nuna asalin URL na asali.
Tantance Ayyukan Agent Kafin Su Gudana: Ƙofar Decision Model ta JevAct — Sebastien Rousseau JevAct na duba zaɓaɓɓun ayyukan LLM agents kafin su gudana. Kan ReAct mai LLM ɗaya, riba ta daidaita a Telecom da BFCL amma ta gauraya a Airline. Originally published at https://sebastienrousseau.com/ha/2026-11-18-tantance-ayyukan-agent-kafin-su-gudana/ by Sebastien Rousseau. Licensed under CC-BY-4.0.
