Agensh Uyasebenza

I-Agensh iwuhlaka lokusebenzisana kwama-agent amaningi olwakhiwe yi-Microsoft Research. Wonke ama-agent alandela umjikelezo ofanayo wezinyathelo ezinhlanu wokusebenzisana, ezihlela ngokwazo ukwaba umsebenzi zisebenzisa ikhosombe lekhodi elabiwe, imiyalezo esheshayo kanye nebhodi elihlelekile. Ukuhlola kubonisa ukuthi, ngaphandle kokushintsha imodeli, ukwandisa inani lama-agent kusuka ku-1 kuya ku-1024 kukhuphule izinga lokuphasa kwezivivinyo zokuphindaphinda isoftware lisuka ku-33.89% laya ku-55.06%, okufakazela ukuthi inani lama-agent ngokwalo lingaba yisici esisha sokwandisa amakhono e-AI.

Ukuhlelwa okungenamaphakathi: Kususwa i-agent ephethe, wonke ama-agent azithathela imisebenzi, axoxisane ngokwabiwa komsebenzi futhi ahlanganise imiphumela.
Umjikelezo wokusebenzisana wezinyathelo ezinhlanu: I-agent ngayinye iphindaphinda umjikelezo we-asynchronous othi ”qoqa umongo→thatha umsebenzi omncane→wenza futhi wabelane→qinisekisa imiphumela→hlanganisa inqubekelaphambili”.
Indawo yokusebenza eyabiwe: Ikhodi ilawulwa nge-Git (Gitea), kusekelwa ukuthuthukiswa ngesikhathi esifanayo, ukuhlanganiswa kwama-branch kanye nokuboniswa kwezingxabano.
Ukuxoxisana ngemiyalezo: I-Mattermost isetshenziselwa izimemezelo zeqembu nemiyalezo eyimfihlo yomuntu nomuntu; imiyalezo eyimfihlo ingaphazamisa omunye ngokushesha ukuze kuxazululwe ukushayisana kwemisebenzi.
Ibhodi lomongo elabiwe: Liqopha izinhlobo ezinhlanu zezinto, i-OBSERVED/FACT/FAIL/CLAIM/PATCH_SUMMARY, ukuze wonke umuntu aphinde asebenzise ulwazi futhi agweme umsebenzi ophindwayo.
Ukwandiswa kufika ezingeni lezinkulungwane: Ekuhlolweni, inani lama-agent lakhushulelwa ku-1,024, kodwa ukusebenzisana okuzihlelayo kwaqhubeka ngaphandle kokuwa.
Ukuvela okuzenzakalelayo kwezindima: Lapho inhlangano ikhula, ukwabiwa kwemisebenzi okufana nemikhuba yokubuyekeza, izinqubo ezijwayelekile nabahlanganisi abakhethekile kuvela ngokwako ngaphandle kokusethwa kusengaphambili.

Isakhiwo esinezendlalelo esihlukanisa ungqimba lwenhlangano nongqimba lwe-agent eyodwa: I-Agensh ayithathi indawo yohlaka oluyisisekelo lwe-agent eyodwa; yengeza ungqimba lwenhlangano phezu kwalo. Umjikelezo wendawo we-agent ngayinye uphathwa yi-harness eyisisekelo, kanti i-Agensh iphethe kuphela ubunikazi bama-agent, izivumelwano zokusebenzisana, ukuhanjiswa kwemicimbi, izinsizakalo ezabiwe kanye nokubuyiselwa ekuphazamisekeni. Izixhumanisi phakathi kwalezi zingqimba ezimbili zigcinwa zilula futhi zixhunywa kalula. Umjikelezo wokusebenzisana awufakwanga ngokuqinile ku-runtime, kodwa wenziwa ngemiyalelo yokugeleza komsebenzi ekwi-prompt. Wonke ama-agent athola imiyalelo efanayo ngokuphelele; ukuvumelanisa nohlaka oluhlukile oluyisisekelo kudinga kuphela ukushintshwa kwe-adapter elula, ngaphandle kokuguqula umqondo wokusebenzisana.
I-runtime ye-asynchronous eqhutshwa yimicimbi: I-agent ngayinye iqhutshwa yimicimbi ehambisanayo, hhayi umhleli ophakathi. I-router ilalela ngokuqhubekayo imisebenzi yekhosombe le-Gitea nemiyalezo ye-Mattermost, ibhale imicimbi emisha emgqeni oqhubekayo ogcina idatha futhi ususa okuphindiwe, one-ID ehlukile. I-dispatcher iyayihlanganisa, iyihlele bese iyifaka ekujikelezweni okulandelayo kwe-agent. Imiyalezo ilethwa ngamazinga amabili okubaluleka: imiyalezo yeziteshi zomphakathi ifika ekuqaleni komjikelezo, kanti imiyalezo eqondile nezinto ezintsha zebhodi zifakwa emiphumeleni yezicelo zamathuluzi zomjikelezo wamanje ukuze zifike phakathi kwawo. Lokhu kwenza i-agent ikwazi ukuphazamisa ozakwabo ngesikhathi sangempela lapho imisebenzi ishayisana. Izindlela ezihambisanayo zihlanganisa ukuphinda uzame kancane kancane ngemva kokwehluleka, ukubuyiselwa komugqa nemihlangano ngemva kokuqalisa kabusha i-router, ukuvusa ngokuzenzakalelayo ama-agent angasebenzi imizuzu engu-10, kanye nesikhumbuzi sokuthumela into yebhodi uma kuqedwe umjikelezo ngaphandle kokuthumela lutho.
Umongo owabiwe onezinhlobo ezithile wakha inkumbulo yenhlangano: Ngokulandela umqondo we-DeLM, i-Agensh ibeka inkumbulo yenhlangano ngaphandle kwe-agent njenge-log ehlelekile ye-append-only. Izinto zihlukaniswe ngezinhlobo ezinhlanu: i-OBSERVED iqopha ukuziphatha okubonakalayo kohlelo oluyisithenjwa, i-FACT iqopha amaqiniso aqinisekisiwe, i-FAIL iqopha imibono eqinisekiswe njengengalungile, i-CLAIM isetshenziselwa ukumemezela indawo ethathwayo ukuze kugwenywe ukushayisana kwemisebenzi, kanti i-PATCH_SUMMARY ifingqa izinguquko eziqediwe ngefomethi engaguquki ethi ”files= | idea= | evidence=”. Ibhodi lisatshalaliswa ngezindlela ezimbili: ekuqaleni komjikelezo omusha kufakwa amarekhodi akamuva angu-2000, kanti phakathi komjikelezo izinto ezintsha zengezwa ngemiphumela yezicelo zamathuluzi. Kukhona futhi ithuluzi le-board_grep elisekela ukusesha amagama angukhiye e-AND/OR kuwo wonke umlando, lidlule umkhawulo wewindi lomongo le-agent eyodwa futhi lisekele ukuqoqwa nokusetshenziswa kabusha kolwazi emisebenzini emide.
Umjikelezo wokusebenzisana unendlela eyakhelwe ngaphakathi yokuxazulula izingxabano nokuqinisekisa: Lapho ithatha umsebenzi omncane, i-agent kufanele ishicilele isimemezelo se-CLAIM ebhodini esichaza indawo yayo. Uma ithola ukugqagqana nomunye, bobabili baxoxisana ngemiyalezo eyimfihlo ukuze balungise ukwabiwa. Esigabeni sokuhlanganisa, i-Git iyisisekelo: i-agent ithuthukisa ku-branch yangasese bese ithumela i-PR ku-main branch. Uma kuvela ingxabano yokuhlanganisa, kufanele ilande inqubekelaphambili yakamuva, ixazulule ingxabano bese iphinda ithumele, futhi kufanele iqale iqinisekise umphumela wendawo ngokumelene nemigomo yokwamukelwa ngaphambi kokuhlanganisa. Le ndlela ihlukanisa imisebenzi emithathu evame ukwenziwa umphathi, okuwukwaba imisebenzi, ukulamula izingxabano nokuhlanganisa imiphumela, iye emjikelezweni wansuku zonke wawo wonke amalungu, futhi yenze umsebenzi wokuhlanganisa ubonakale futhi ulandeleke ngomlando wezinguqulo nenqubo ye-PR.

Ukulungiselela indawo: Qinisekisa ukuthi umshini une-Docker kanye ne-API yemodeli esebenzayo.
Ukuklona ikhosombe lekhodi: Landa ikhosombe le-microsoft/Agensh ku-GitHub uye endaweni yangakini, bese uliklona lapho ikhodi isivulelekile.
Ukuphakela ingqalasizinda eyabiwe: Qalisa izinhlobo ezintathu zezingxenye ezidingwa uhlaka, i-Gitea (ikhosombe lekhodi), i-Mattermost (ukuxhumana ngemiyalezo) kanye nensizakalo yebhodi lomongo owabiwe.
Ukulungisa amapharamitha ama-agent: Ekucushweni, setha inani lama-worker, umgomo womsebenzi, isabelomali sesikhathi, kanye nohlobo lwe-harness eyisisekelo ye-agent eyodwa nokhiye wemodeli.
Ukuqalisa inhlangano: Sebenzisa iskripthi sokuqalisa; uhlaka luzovula ama-agent ngamaqoqo ngokohlelo olunezikhathi ezihlukene, kanti ukuphakelwa okukhulu kuzosatshalaliswa ngokuzenzakalelayo ezindaweni eziningi.
Ukusebenza ngokuzimela kwama-agent: Amalungu azithathela imisebenzi, athuthukise ikhodi, athumele ama-PR futhi ahlanganise ku-main branch ngokomjikelezo wezinyathelo ezinhlanu, ngaphandle kokungenelela komuntu. Okudingekayo kuphela ukubuka inqubekelaphambili esiteshini somphakathi se-Mattermost.
Ukuvala nokuletha: Lapho isikhathi sokugcina sisondela, uhlaka luzokhumbuza ngokuzenzakalelayo ukuthi umsebenzi uqoqwe. Umphumela wokugcina ulethwa ngekhodi ye-main branch namarekhodi ebhodi.

Ukudlula umkhawulo womphathi: Ngemva kokususa umhleli ophakathi, ukunwebeka kohlelo akusaboshiwe umkhawulo wekhono le-agent eyodwa lokuxhumanisa, futhi lungakhula kalula lufike kuma-agent angu-1,024.
Inani liba yisici esisha sekhono: Ngaphandle kokushintsha imodeli noma uhlaka oluyisisekelo, ukwandisa kuphela inani lama-agent kungakhuphula isilinganiso sokuphasa kwezivivinyo cishe ngo-49%, kuvule indlela yesibili ye-scaling ngaphandle kobukhulu bemodeli.
Ukunciphisa kakhulu ukubambezeleka komsebenzi: Ama-agent angu-128 afinyelela izinga lokuphasa elingaphezu kuka-30% ngomzuzu wama-30, kanti i-agent eyodwa ayizange ifinyelele ezingeni elidingekayo emahoreni amabili okuqala. Iqembu elikhulu lingafinyelela umphumela ofanayo ngesikhathi esincane.
Ingqalasizinda ilula futhi ihlukanisiwe: Umjikelezo wokusebenzisana wenziwa ngemiyalelo ye-prompt, hhayi ngekhodi eqinile, ngakho ungahambisana kalula namafreyimu ayisisekelo ahlukene afana ne-Copilot ne-Claude Code. Izinsizakalo ezabiwe zidinga kuphela izingxenye ezintathu zomthombo ovulekile, i-Gitea, i-Mattermost nebhodi.
Ulwazi lwenhlangano lungagcinwa luphinde lusetshenziswe: Ibhodi elinezinhlobo ezithile lihlela okuboniwe, amaqiniso nemibono ehlulekile. Izinto ze-FAIL ikakhulukazi zivimbela ozakwabo ekuchitheni isabelomali ngokuphinda umsebenzi ofanayo, kwakhe inkumbulo yezinga lenhlangano.

Iwebhusayithi yephrojekthi：https://agens-harness.github.io/project/
Iphepha lobuchwepheshe le-arXiv：https://arxiv.org/pdf/2609.26781

Ukuphindaphinda ukuziphatha kwesoftware enkulu: Ezimweni lapho kungekho khona uxhumano lwenethiwekhi futhi isikhathi silinganiselwe, kusetshenziswa amakhulu kuya ezinkulungwaneni zama-agent ukwakha kabusha kusukela ekuqaleni ukuziphatha kwesoftware enjenge-FFmpeg ne-pandoc enamakhulu ezinkulungwane zemigqa, bese kudluliswa izivivinyo ezifihliwe.
Ukuthuthukiswa okuphuthumayo okudinga isikhathi: Kufanele imisebenzi enesabelomali sesikhathi esiqinile. Ama-agent angu-128 angafinyelela ngemizuzu engu-30 izinga lokuqedwa i-agent eyodwa engakwazi ukulifinyelela emahoreni amabili.
Ukuhlolwa kwekhodi ngezinga elikhulu kakhulu: Ama-agent athatha amamojula ngokwehlukana, abhale izivivinyo ngesikhathi esifanayo futhi abuyekezane ama-PR, okwenza afaneleke ngokwemvelo ekuhloleni ukuphepha nasekukhuliseni ukumbozwa kwezivivinyo kumakhosombe amakhulu ezinkampani.
Ucwaningo nentuthuko yokuhlola ngaphandle kohlelo olubekwe kusengaphambili: Lapho umsebenzi ungeke uhlukaniswe kusengaphambili, ama-agent athola imisebenzi ngokwawo ngokubheka ukuziphatha okuyisithenjwa. Lokhu kufanelekile ekuthuthukisweni kobuchwepheshe lapho izidingo zingacacile futhi imingcele inqunywa ngokuhlola.
Ukuthuthwa kwesoftware eyisisekelo phakathi kwemikhakha: Ekuthuthukiseni amasistimu emikhakheni ehlukahlukene, njengabahumushi be-PHP nokulingisa ama-molecule, ubuhlakani beqembu bungamboza ulwazi olungatholakali ku-”genius” oyedwa, kanti ochwepheshe bomkhakha ofanayo bangasekelana futhi bathathe izintambo.