Kwenzekeni
I-36Kr ibika ukuthi i-OpenAI ikhiphe okuthunyelwe kukanjiniyela ngo-Agasti 20 echaza i-Codex njengenkundla eyakhelwe phezu kwehhanisi lomenzeli ovulekile. Ngokombiko, abathuthukisi bangasebenzisa i-codex exec, i-SDK noma i-app-server ukuze bashumeke amakhono e-Codex kuma-consoles ebhizinisi, izinhlelo zesevisi yamakhasimende, amathuluzi okuvikela nezinhlelo zokusebenza zangaphakathi. Umbiko wethula lokhu njengokusuka kokuthi Codex ekubeni ngokuyinhloko umsizi wohlelo ukuze isebenze njengengqalasizinda ephinde isetshenziswe.
I-36Kr ibika ukuthi okuthunyelwe kukanjiniyela we-OpenAI ka-Agasti 20 wethule i-Codex njengenkundla futhi kwachaza i-Codex Harness enomthombo ovulekile ngaphansi kwe-Codex App, i-CLI nezandiso ze-IDE. Umbiko uthi ihhanisi lilawula isimo seseshini, umongo, izingcingo zamathuluzi, ama-sandbox kanye nokugunyazwa kwabantu. Iphinde ithi abathuthukisi bangaxhuma lawo makhono emikhiqizweni ekhona ngokusebenzisa i-codex exec, i-SDK noma iseva yohlelo lokusebenza esekela imicu, ukuphenduka, ukusakazwa kwemicimbi kanye nezivumelwano zokugunyaza. Awekho umbhalo oyinhloko we-OpenAI onikezwe umthombo, ngakho le mininingwane ifakwe ku-36Kr futhi ayiqinisekiswanga ngokuzimela lapha.
Umbiko uthi amakhono ethulwa ngezigaba. Ichaza ukutholakala kwe-Codex SDK kanye nokukhishwa okuvamile kwe-Codex ngo-Okthoba 2025, ukusetshenziswa kamuva kwe-codex exec yemibhalo nemisebenzi eqhubekayo yokuhlanganisa, kanye nokusekelwa kweseva yohlelo lokusebenza kumaseshini aqhuba isikhathi eside. Iphinde ithi i-OpenAI ishicilele incazelo ye-loop ye-ejenti ye-Codex ngoJanuwari 2026, ihlanganisa ukuthi imodeli ithola kanjani imiyalo, amathuluzi ezingcingo, ifunda imiphumela futhi iqhubekele esinyathelweni esilandelayo. Incazelo emaphakathi ye-36Kr ukuthi i-OpenAI manje isihlanganise le misebenzi ngaphansi kwegama elithi Codex Harness futhi ikhuthaza abathuthukisi ukuthi bakhe imikhiqizo eduze kwayo.
Ngokusho kwe-36Kr, i-OpenAI ikhombe izinhlelo zokusebenza ezimbalwa kokuthunyelwe kwayo kwebhulogi. Umbiko uthi i-Cisco isebenzisa i-Codex SDK ku-App Builder ye-Cisco Cloud Control, kuyilapho i-GitHub ne-JetBrains ihlanganisa i-Codex ezindaweni ezikhona zokuthuthukiswa. Iphinde ithi i-Thrive Holdings ne-Crete zisebenzisa i-Codex ukulungiselela intela nokuthi umshayeli wendiza oyedwa uphathe izimbuyiselo zentela ezingu-7,000 kuyilapho enciphisa isikhathi sokulungiselela cishe ngengxenye eyodwa kwezintathu. Lezi zibonelo nezilinganiso zethulwa ngu-36Kr njengezimangalo ezivela ku-OpenAI; umthombo awuhlinzeki ngokuhlolwa okuzimele, indlela yokusebenza noma imibhalo evela ezinhlanganweni ezihililekile.
Umbiko uqhathanisa i-Codex Harness ne-DeepSeek Harness, othi DeepSeek ivuleke ngo-Agasti 14 ngaphansi kwesifinyezo se-DSH. I-36Kr ithi i-DSH iphatha amamodeli, amathuluzi, amakhono, izikhathi, ama-sandbox, isitoreji, iluphu ye-ejenti, ukuhlela kanye nesixhumi esibonakalayo somsebenzisi njengama-plugin aphethwe ngesistimu ebizwa nge-Cordis. Igqamisa i-Codex njengesikhathi sokusebenza esididiyelwe kakhudlwana izingxenye zayo ezingezona ezibalulekile okufanele zishintshwe enjinini yayo, kuyilapho ichaza i-DSH njengemodular kakhudlwana. I-athikili iphinde idingide i-Kimi Code kanye ne-ZCode njengamasistimu e-ejenti ahlobene kodwa ahlelwe ngendlela ehlukile, kunokuzethula njengomkhiqizo ofanayo.
Imininingwane yomthombo: eu.36kr.com ↗
Kungani kubalulekile
Ushintsho olubikiwe lunganikeza i-OpenAI indima enkulu kusendlalelo sesofthiwe esinquma ukuthi ama-agent e-AI awasebenzisa kanjani amathuluzi, alondoloze umongo, acele ukugunyazwa futhi aqedele umsebenzi wezinyathelo eziningi. Leso sendlalelo singathinta ukwethembeka, izindleko, ukubonakala nokulawulwa komsebenzisi ngaphandle kwemodeli eyisisekelo. Umbiko uphinde ubeke umnyakazo we-OpenAI emqhudelwaneni nesikhathi sokusebenza se-ejenti evulekile, okuhlanganisa i-DeepSeek Harness namanye amaphrojekthi omthombo ovulekile.
Umnyakazo obikiwe ubalulekile ngoba isikhathi sokusebenza somenzeli silawula izinyathelo zokusebenza phakathi kokukhiphayo kwemodeli nomsebenzi oqediwe. Ngokusho kwe-36Kr, i-Codex Harness ingakwazi ukugcina umongo, amathuluzi okushaya ucingo, isebenzise amabhokisi esihlabathi kanye nezenzo zomzila ngokusebenzisa izimvume zomuntu. Ezinhlanganweni ezishumeka ama-ejenti kusofthiwe yangaphakathi, leyo misebenzi inganquma ukuthi i-ejenti iyabonakala, iyaphazamiseka futhi iyahambisana nezimvume ezikhona. Umbiko awuqinisekisi ukuthi i-Codex Harness ihlangabezana nanoma yikuphi ukuphepha kwebhizinisi okuthile noma izinga lokuthobela.
I-36Kr ibika ukuthi i-OpenAI yethule ukuqhathanisa kwe-ARC-AGI-3 lapho i-GPT-5.6 Sol ithole u-13.3% iyodwa kanye no-38.3% nge-Codex Harness yokucabanga eqhubekayo namandla okucindezela umongo. Lesi sihloko siphinde sithi umthamo wethokheni lokukhipha wehlele cishe ingxenye eyodwa kweziyisithupha yenani lokuqala. Lezo zibalo, uma zikhiqizwa kabusha, zingaphakamisa ukuthi i-orchestration nokuphathwa komongo kungashintsha kakhulu ukusebenza nezindleko zemodeli efanayo. Kodwa-ke, umthombo awuhlinzeki ngephrothokholi yokuhlola, imininingwane eyisisekelo, ukuphindaphinda okuzimele noma ulwazi mayelana nokubaluleka kwezibalo, ngakho izinombolo kufanele zithathwe njengemiphumela yenkampani ebikiwe esikhundleni sobufakazi obuqediwe.
I-athikili iphikisa ukuthi izinkampani zamamodeli zinezisusa zokuba ngabanikazi noma zokusabalalisa isikhathi sokusebenza kumamodeli azo. I-36Kr ithi isikhathi sokusebenza singadalula lapho imisebenzi yehluleka khona, lapho ithuluzi libiza khona, ukuthi ukuzama futhi ukwenzeka kaningi kangakanani nokuthi yimaphi amasu omongo adla amathokheni ambalwa. Lokhu kungasiza umthuthukisi wemodeli athuthukise kokubili amamodeli akhe kanye nesistimu yakhe yokusebenza. Umthombo uphinde uphawule ukuthi noma yikuphi ukusetshenziswa kwamarekhodi omsebenzi wokuqeqeshwa kuzoncika kuzinqubomgomo zobumfihlo ezisebenzayo; ayiqinisekisi ukuthi i-Codex noma i-DeepSeek iqoqani, igcine ini noma isebenzise ini.
Isikhathi sokusebenza esivuleke kakhulu singaphinda siguqule ibhalansi phakathi kwabathengisi bemodeli nabathuthukisi bohlelo lokusebenza. I-36Kr ithi i-plugin-based plugin ye-DSH iklanyelwe ukuvumela abathuthukisi ukuthi bathathele amamodeli, amathuluzi, isitoreji, amabhokisi esandbox, nezingxenye zeluphu ngaphandle kokugcina imfoloko ehlala isikhathi eside yayo yonke iphrojekthi. Lokho kuvumelana nezimo kungase kukhange amaqembu afuna ukushintsha abahlinzeki noma amasu okusebenzisa. Ngesikhathi esifanayo, i-athikili ibika ukuthi i-DSH isewukubuka kuqala futhi ukuthi impendulo yomphakathi iphakamise ukukhathazeka mayelana nesivinini, ukusetshenziswa kwamathokheni, imibhalo, ukuhambisana kanye nekhwalithi ye-plugin engalingani. Lokho kubuka kuyimpendulo, hhayi ukuhlola okuhlelekile.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
Ongakubuka ngokulandelayo
Imibuzo ebaluleke kakhulu evulekile ukuthi i-Codex Harness isetshenziswa ngempela yini ngaphandle kwemodeli ye-ecosystem ye-OpenAI, iziphi izingxenye zonjiniyela abangazifaka esikhundleni, ukuthi idatha evela ekuhambeni komsebenzi okushunyekiwe isingathwa kanjani, kanye nokuthi izinzuzo zokusebenza okubikiwe zihlala ngaphansi kokuhlolwa okuzimele. Umbiko onikeziwe awuqinisekisi ukutholakala okubanzi, amanani entengo, imigomo yenkontileka, izinqubo zobumfihlo noma ukwethembeka kokukhiqiza.
Okokuqala, bheka ubufakazi bokuthi i-Codex Harness ingasebenza ngamamodeli namathuluzi angaphezu kwesitaki esithandwayo se-OpenAI. I-36Kr ichaza i-Codex njengevuliwe futhi ishumekeke, kodwa umbiko onikeziwe awucacisi ilayisensi, izingxenye ezishintshayo ngqo, ama-adaptha amamodeli asekelwayo, izinketho zokuphakela noma ukuthi izingxenye ezibalulekile zihlala zivaliwe. Leyo mininingwane izonquma ukuthi onjiniyela bayasithola yini isikhathi sokusebenza esiphathekayo noma ukuhlanganiswa kwe-OpenAI okuhlanganiswe ngokuqinile.
Okwesibili, ukuhlolwa okuzimele kufanele kuhlole ukuthuthukiswa okubikiwe kwe-ARC-AGI-3 kanye nokunciphisa amathokheni. Ukuqinisekisa okuwusizo kuzobandakanya izinguqulo eziqondile zamamodeli, ukwaziswa, izilungiselelo zamahhanisi, imikhawulo yokuqukethwe, ukulungiselelwa kwamathuluzi, ukuphatha ukwehluleka kanye nokuqagela kwezindleko. Ngaphandle kwaleyo mininingwane, ukuqhathanisa akukwazi ukukhombisa ukuthi inzuzo iqhamuka ekwakhiweni kwesikhathi sokusebenza okuvamise ukusebenziseka noma ekusetheni okuthile kobunikazi.
Okwesithathu, izinhlangano ezicabangela ama-ejenti ashumekiwe zizodinga ulwazi olucacile mayelana nokuphathwa nokulawulwa kwedatha. Umbiko uchaza ikhodi yebhizinisi, idatha yekhasimende, amathuluzi angaphakathi, ukugunyazwa namarekhodi emisebenzi ahamba ngesikhathi esabiwe. Ayisho ukuthi lawo marekhodi acutshungulwa kuphi, agcinwa isikhathi esingakanani, ukuthi abalawuli bangakwazi yini ukuwahlola noma ukuwasusa, noma ukuthi angase asetshenziselwe ukuthuthukisa imodeli. Lokho okungaziwa kuyimpahla yokuthunyelwa okubandakanya umsebenzi oyimfihlo noma olawulwayo.
Okokugcina, landelela ukuthi ingabe iqoqo elikhulayo lezikhathi zokusebenza zomenzeli liba yimakethe yengqalasizinda eqinile noma lihlala liyisethi yamathuluzi onjiniyela aqondene nemodeli ethile. I-36Kr ibika ukuthi i-DeepSeek Harness idonse ukunaka okukhulu kwangaphambi kwesikhathi konjiniyela nokuthi i-Kimi Code ne-ZCode sezivele zinikeza amandla okusebenzisa ahlobene. Ukwamukelwa ekukhiqizeni kuzoncika ekuzinzeni, ezindleleni zokubuyisela emuva, imingcele yemvume, imibhalo, ukubikezelwa kwezindleko kanye nokubuyekezwa kokuphepha okuzimele. Umbiko onikeziwe usungula inzalo kanye nokuma komkhiqizo, kodwa hhayi ukutholwa kwebhizinisi okubanzi noma ukuphakama okuthembekile.