Kwenzekeni
Iphrinta ethunyelwe ku-arXiv ngomhla zingama-29 kuJulayi 2026 futhi yabuyekezwa ngomhlaka-7 Agasti iphakamisa indlela entsha yokukala inqubekelaphambili ocwaningweni oluzenzakalelayo lwe-AI: nikeza i-ejenti umbuzo ovulekile omaphakathi wephepha elingashicilelwe lekhwalithi ephezulu futhi wenze ababhali balelo phepha balinganise umphumela. Ezivivinyweni ezimbili, ama-ejenti asemngceleni aqedele ubunjiniyela ngaphandle kosizo lomuntu kodwa enqatshwa abafundi bebanga labantu, futhi ababhali bakhomba izindlela ezinhlanu zokuhluleka eziphindelelayo.
Ukuphrinta kwangaphambili ku-arXiv, okuthunyelwe zingama-29 kuJulayi 2026 futhi kwabuyekezwa ziyisi-7 Agasti 2026, kubuza ukuthi ama-ejenti e-AI angakwazi yini ukwenza ucwaningo lwe-AI oluvulekile. Iphethe ababhali abasohlwini abangu-24, okuhlanganisa uPeter Kirgis, Sayash Kapoor, Helen Toner, Gillian Hadfield, Seth Lazar, Rishi Bommasani kanye no-Arvind Narayanan, futhi ifakwe ngaphansi kobuhlakani bokwenziwa, amakhompyutha nomphakathi, nokufunda ngomshini. Umnikelo wawo omaphakathi uyindlela yokulinganisa ababhali abayibiza ngokuthi "ukuhlola ithunzi": i-ejenti inikezwa umbuzo wocwaningo omaphakathi, ovulekile wephepha lekhwalithi ephezulu kodwa elingashicilelwe, bese ababhali bangempela bephepha bese bebeka amagremu lokho okukhiqizwa yi-ejenti.
Ababhali babeka lokhu njengenketho yesithathu phakathi kwezindlela ezimbili ezikhona abazichaza njengezingenele. Ukuhlola kwamanje ama-ejenti, bayabhala, noma ama-ejenti okuhlola emisebenzini emincane, eqinisekiswayo - okuthi ngokwakha kungafaki ucwaningo oluvulekile - noma athumele amaphepha akhiqizwe yi-AI ekubuyekezweni kontanga abayizimpumputhe, abawaveza njengawande kakhulu, agxilile futhi ahlushwa yikhwalithi embi yokubuyekeza. Ukuhlola izithunzi esikhundleni salokho kusebenzisa abantu asebeyazi inkinga, izincwadi zayo kanye nezingibe zayo njengabadidiyeli, ngoba banempendulo engashicilelwe yokuqhathanisa nayo.
Ithimba liqhube le ndlela kokuthunyelwe okungakashicilelwa okubili ku-NeurIPS 2026. Ngokusho kwe-abstract, ama-ejenti asemngceleni anikezwe izinsuku eziyisithupha kanye nezinkulungwane zamadola wekhompiyutha umzamo ngamunye. Ama-ejenti aqedele bonke ubunjiniyela ngaphandle kosizo lomuntu - ikhodi, izivivinyo, ingqalasizinda - kodwa abakwazanga ukwenza inqubekelaphambili enkulu ekuphenduleni imibuzo yocwaningo ngokwabo. Womabili amaphepha angumphumela, ngokwamagama ababhali, enqatshwe ngokusobala ababhali bokuqala abasebenza njengababuyekezi.
Kusuka emibhalweni ababhali bakhipha izindlela ezinhlanu zokuhluleka eziphindelelayo: ukwahlulela okubi mayelana nebha yocwaningo olungashicilelwa; izimpendulo ezingelona iqiniso kumaphutha ekwakhiweni kocwaningo; ukuhlehla okungalungile kusuka emikhawulweni efile; ukungaqondi kahle kwezinsiza; nokukhukhuleka kweziyalezo. Babika ukuthi ukuhlola ukuqina kusetshenziswa imodeli yesibili kanye nesikafula sesibili kukhiqize kabusha ukwehluleka okufanayo, okuphikisana nomphumela wokuba i-artifact yesistimu eyodwa noma ihhanisi elithile. Ithimba lithi likhipha izibuyekezo zochwepheshe, izimpendulo zenhlolovo, amakhosombe e-ejenti namalogi.
Izinto ezimbalwa ezingabonakali ezingabonakali kufanele zichazwe ngokucacile. Ayisho ukuthi yimaphi ama-ejenti, amamodeli noma izikafula ezisetshenzisiwe, noma isistimu yesibili ekuhloleni ukuqina. Ayizikhombi ukuthunyelwa kwe-NeurIPS okubili noma izinkundla ezingaphansi kwazo, inikeza inani eliqondile lekhompiyutha elingaphezu "kwezinkulungwane zamadola," ichaza irubhrikhi yokugreda, noma isho ukuthi ingabe abafundi bebanga baphuphuthekiswe iqiniso lokuthi bebebuyekeza ukukhishwa kwe-ejenti. Umsebenzi ungowokuphrinta; akukho lutho emthonjeni okubonisa ukuthi sekuqedile ukubuyekezwa kontanga. Futhi isampula imibuzo emibili yocwaningo, ehlelwe abantu abayibuzile.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Izimpikiswano zokuthi inqubekelaphambili ye-AI izosheshisa kakhulu incike ekutheni izinhlelo ze-AI zingathatha ucwaningo lwe-AI ngokwalo. Lokhu kuwubufakazi obuqondile, uma buncane, bokuthi isisekelo asikabambi, futhi sihlukanisa ama-ejenti wengxenye angakwazi ukukwenza kakade - ukuqaliswa - kusukela ezingxenyeni abangakwazi: ukwahlulela ukuthi yini okufanele ishicilelwe, ukuhlela kabusha isifundo esinephutha, nokushiya iziphetho ezifile.
Uhlaka lwephepha luqonda ngqo ekucabangeni okuthwala umthwalo ekubikezelweni kwamanje kwe-AI. Njengoba i-abstract ikubeka, izibikezelo zenqubekelaphambili ye-AI eqhumayo zincike kubenzeli be-AI abazenzela ucwaningo lwe-AI - iluphu lapho amasistimu angcono akha amasistimu angcono. Leyo loop ivamise ukugonyelwa futhi ayivamisile ukukalwa, ngoba into eyenziwa ngokuzenzakalela iwuhlobo lomsebenzi olumelana nombhalo wokushaya amaphuzu. Indlela ekhiqiza isinqumo esenziwe ngezigaba esivela kubacwaningi abasebenzayo embuzweni obukhoma, ongaphenduliwe iwuphenyo oluqondile lwalokho kuqagela kunamaphuzu ebhentshimakhi emisebenzini enezimpendulo ezaziwayo.
Umphumela udweba umugqa phakathi komsebenzi kunokuwuzungezile. Ama-ejenti aqhube izinhlelo zokuhlola zezinsuku eziningi zinganakiwe futhi akhiqiza ikhodi yokusebenza nemiphumela — lokho amandla angempela kanye noshintsho olunengqondo kusukela eminyakeni embalwa edlule. Abangazange bakwenze ingxenye enqumayo ukuthi ucwaningo lukufanele yini ukushicilelwa: ukuqaphela ukuthi umphumela mncane, ukusungula indlela ezungeze iphutha lomklamo, nokwazi ukuthi umugqa wokuhlasela ufile nini. Emaqenjini anquma ukuthi abeke kuphi ama-ejenti ocwaningweni noma epayipini lobunjiniyela, lowo mehluko uyasebenza kakhulu kunomphumela wamakhono ahlanganisiwe.
Izindlela zokwehluleka ezimbili kwezinhlanu zivame ukudlula ucwaningo. Ukungazi kahle kwezinsizakusebenza - ukushisa ikhompuyutha engaguquki noma isabelomali sesikhathi ngaphandle kokulungisa - kanye nokwehla kweziyalezo, lapho i-ejenti iyeka kancane kancane ukwenza lokho ebiceliwe, kuyizinkinga eziphindaphindayo kunoma yikuphi ukuthunyelwa okuzimele okuhlala isikhathi eside, kusukela kuma-ejenti amakhodi wezinsuku eziningi kuya kokuzenzakalelayo okusebenza. Ukugijima kwezinsuku eziyisithupha lapha kude ngokungavamile ngezindinganiso zokuhlola, futhi ama-horizons amade yilapho kanye lokhu kwehluleka kunendawo yokuhlanganisa ngaphambi kokuba noma ubani abone.
Imikhawulo ingokoqobo futhi ababhali abazifihli. Izibonelo ezimbili zezifundo endaweni eyodwa, ezihlolwe ngesifinyezo esisodwa samamodeli, azikwazi ukusekela isimangalo mayelana nokuthi ama-ejenti azokwazi ukwenzani ngonyaka ozayo. Lona umphumela oyize onesitembu sesikhathi. Ababhali bebanga nabo banesandla esisobala embuzweni abakhethe ukusebenza kuwo, futhi indlela ngokwengxenye iyazidla: uma iphepha selishicilelwe, umbuzo walo wocwaningo ngeke usakwazi ukusebenza njengokuhlola okungabonakali. Ukukhipha izibuyekezo, amakhosombe namalogi yikhona okwenza okutholakele kuhloleke kunesimangalo okufanele sithathwe ngokuthenjwa.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').What is the most accurate way to describe what AI Agents can do today?
Ongakubuka ngokulandelayo
Ukuthi ukuhlolwa kwethunzi kuyaphindwa yini ngesilinganiso esikhulu ngokuhlelwa okungaboniwe, okubhaliswe ngaphambilini; ukuthi ngabe ama-ejenti amasha nama-scaffolds avala igebe lapho kuqhutshwa kabusha amaqoqo namalogi akhishiwe; kanye nokuthi izindlela ezinhlanu zokwehluleka ziyaqhubeka yini ngaphansi kokuhlolwa okuzimele nokubuyekezwa kontanga.
Into yokuqala okufanele uyibukele ukuthi ukuhlola ithunzi kuyaba yini ithuluzi eliphindaphindwayo noma kuhlale kuwumboniso ophuma kanye. Lokho kungasho amaphepha amaningi ezindaweni eziningi nasezinkundleni ezingaphansi, amarubrikhi okugreda abhaliswe ngaphambili avunyelwa ngaphambi kokuthi i-ejenti iqalise, futhi ukuhlelwa kuphuphuthekisiwe ukuze kubonakale ukuthi okukhiphayo kuvela kumenzeli noma kumuntu. Kungasho futhi ukubhekana nenkinga yokuhlinzeka: indlela idinga ukugeleza okuqhubekayo komsebenzi ongashicilelwe wekhwalithi ephezulu ababhali bawo abazimisele ukuchitha umzamo wangempela wokubuyekeza emzamweni womenzeli.
Okwesibili ukuthi igebe liyavaleka, futhi ngokushesha kangakanani. Ngenxa yokuthi amakhosombe e-ejenti namalogi ayakhululwa, amanye amaqembu angaphinda aqalise amamodeli amasha nezikafula ngokumelene nemibuzo emibili yocwaningo efanayo futhi aqhathanise ngokuqondile. Bheka ukuthi ukuthuthukiswa kuvela yini njengonjiniyela ongcono - lokho ama-ejenti asevele akwenzile - noma njengokunyakaza kokwehluleka kokwahlulela, okuyingxenye iphepha elithi ayikho. Ukuphinda kuphindwe okukhiqiza ukuhlola okuningi kodwa ukwenqatshwa okufanayo kungaba isignali ehlukile kunaleyo ekhiqiza iphepha ababhali bangempela abalithatha ngokungathi sína.
Okwesithathu, ukucutshungulwa okuzimele kwama-artifact akhululiwe. Izindlela ezinhlanu zokwehluleka ukufunda kwababhali imibhalo yabo; abanye abacwaningi abafunda izingodo ezifanayo bangase bahlukanise ukwehlukana ngokwezigaba, noma bathole ukuthi okunye ukwehluleka kulandelela ku-scaffold kanye nokwazisa kunokuba kumamodeli. Ukuthi ukuphrinta kwangaphambili kuyakusula yini ukubuyekezwa kontanga, nokuthi i-taxonomy yayo iyasinda yini ekuhlolweni kwabanye abantu, izonquma ukuthi ingathwala isisindo esingakanani.
Okokugcina, thatha ngaphandle komphakathi wocwaningo. Uhlu lwababhali luhlanganisa abantu abasebenza ngokuqondile ekubuseni kwe-AI kanye nenqubomgomo, kanye nezikhathi zesikhathi sokuphakelayo kwe-AI R&D okuzenzakalelayo kube yizinhlaka zokuphepha emngceleni, imikhawulo yamandla kanye nezinkulumompikiswano zenqubomgomo kazwelonke. Buka ukuthi lobu bufakazi bucashuniwe yini kulezo zilungiselelo, nokuthi kubikwa ngokunembile yini: iphepha lithi ama-ejenti anamuhla alwa nezingxenye ezibalulekile zomjikelezo wokuphila wocwaningo, ngokusekelwe ezimweni ezimbili - hhayi ukuthi ngeke bakwazi ukwenza ucwaningo, futhi hhayi ukuthi amandla onjiniyela aboniswayo ayinto encane.