Orchard – Uhlaka lwe-Agentic AI oluvulelekile olwakhiwe yi-Microsoft Research Uyasebenza

I-Orchard iwuhlaka lwe-Agentic AI oluvulelekile olwakhiwe yi-Microsoft Research. Umongo walo kusevisi yezindawo i-Orchard Env esekelwe ku-Kubernetes, esekela ukusetshenziswa kabusha kwama-sandbox ezizindeni ezahlukene ukuze kwenziwe i-data distillation, i-reinforcement learning rollout nokuhlola. Uhlaka seluhlanganisa ubunjiniyela besoftware, ukuzulazula ku-browser kanye nomsizi womuntu siqu, futhi lwedlulise idatha yokuqeqesha nezindlela zokuhlola njenge-open source.

Isevisi yezindawo ehlanganisiwe ezizindeni ezahlukene: I-Orchard Env inikeza ukuphathwa komjikelezo wokuphila wama-sandbox, ukwenziwa kwemiyalo, i-file I/O, izinqubomgomo zenethiwekhi kanye ne-REST API. Le ngqalasizinda eyodwa ingasekela ikhodi, amawebhusayithi, ama-desktop, ama-mobile kanye nama-workflow okukhiqiza.
Ama-recipe okuqeqesha ezizindeni ezintathu: I-Orchard-SWE igxile ekulungiseni ikhodi, i-Orchard-GUI isebenza ekuzulazuleni kwamakhasi ewebhu ngokubona, kanti i-Orchard-Claw ibhekene nemisebenzi yansuku zonke yokukhiqiza efana ne-imeyili nekhalenda. Zonke zinikeza amapayipi aphelele okuqeqesha e-SFT + RL.
Ukuqeqesha ngaphakathi kwe-Harness yangempela: Isekela ukuqeqesha nokuhlola okuphelele ngokuqondile kuma-harness asetshenziswayo wangempela afana ne-Codex, OpenClaw ne-ZeroClaw, iqede ukungahambisani phakathi kwendawo yokuqeqesha nendawo yokusetshenziswa.
Amasethi edatha avulekile nezivumelwano zokuhlola: Kukhululwa ama-trajectory angu-107K e-SWE nedatha engu-3,070 ye-GUI multimodal rollout, kanye namabhentshimakhi okuhlola nama-recipe angaphinde akhiqizwe.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokushintshisana ngamaphrojekthi e-AI avulekile
Ungqimba lwendawo oluyi-Kubernetes-native: I-Orchard Env isebenza njengesevisi ezimele, isebenzisa ukuhlelwa kwama-container ukudala nokususa izinkulungwane zama-sandbox ahlukene ngokufanayo. Ukubambezeleka okumaphakathi kokwenziwa komyalo kungamasekhondi angu-0.28 kuphela, futhi kusekela i-auto-scaling ne-Redis distributed lock.
I-Credit-Assignment SFT: Emkhakheni wobunjiniyela besoftware, uhlelo alulahli ama-trajectory ahlulekile ngokwengxenye; kunalokho lukhipha amasiginali okugadwa ezingxenyeni ezisebenzayo, lukhulise inani ledatha yokuqeqesha esebenzisekayo.
I-Balanced Adaptive Rollout + imivuzo eminyene: Ukuze kubhekwane nempendulo encane ye-RL, kusetshenziswa i-balanced adaptive rollout ukubamba amasiginali empumelelo ayivelakancane, bese kwengezwa i-on-policy distillation nemodeli yomvuzo wenqubo esekelwe emithethweni ukuze kunikezwe isiqondiso esinemininingwane ezinyathelweni eziphakathi.
Ukuhlelwa kabusha nge-Value Model: Kusetshenziswa ama-trajectory e-rollout avela ekuhlolweni okungokomlando okungu-20 ukuqeqesha imodeli yenani enamapharamitha angu-4B. Esigabeni sokucabanga, imodeli ikala izixazululo eziningi eziphakanyisiwe ikhethe esingcono kakhulu, okukhuphule i-Orchard-SWE isuka ku-69.7% yaya ku-73.0%.
Amarekhodi e-Harness-Agnostic Agent: I-harness elula yokuqopha ilandelela izingcingo zemodeli ezenziwa yi-harness uqobo bese izihlela kabusha zibe amasampula okuqeqesha. Lokhu kwenza noma iyiphi i-RL codebase ixhumane nanoma iyiphi i-harness ngaphandle kokushintsha i-environment image.

Faka i-SDK: Sebenzisa i-pip install -e "orchard_env[dev]" ukufaka i-Python SDK, bese ulungisa ama-environment variable e-SANDBOX_BASE_URL ne-SANDBOX_API_KEY.
Shayela i-sandbox ngokushesha: Sebenzisa i-SandboxClient context manager ukudala i-sandbox nokusebenzisa imiyalo. Isekela izindlela ezivumelanayo nezingavumelani, ukufunda nokubhala amafayela, ukusebenzisa i-git patch kanye nama-PTY session.
Sebenzisa i-orchestrator: Ngemibhalo emine enikezwa yi-Azure AKS (provision, build, deploy, smoke-test), ukusethwa kwe-orchestrator enama-replica amaningi kungaqedwa cishe emizuzwini engu-20. Kusekelwa nama-cluster angewona awe-Azure.
Qalisa ama-recipe okuqeqesha: Landela ifolda ye-trainer/slime/ endaweni yokugcina ye-GitHub ukuze usebenzise izinqubo eziphelele zokuqeqesha e-SFT + RL ze-Orchard-SWE, Orchard-GUI noma Orchard-Claw ku-Orchard Env.

Amamodeli amancane asondela ekusebenzeni kwamamodeli amakhulu: I-Orchard-SWE ifinyelela ku-73.0% ku-SWE-bench Verified isebenzisa cishe amapharamitha asebenzayo angu-3B, isondela ezinhlelweni ezihamba phambili ezinenani lamapharamitha eliphindwe ngaphezu kuka-10; i-Orchard-GUI ifinyelela esilinganisweni esingu-68.4% kumabhentshimakhi amathathu amakhulu ewebhu ngamapharamitha angu-4B, okucishe kufane ne-OpenAI CUA ne-Gemini CUA.
Izindleko zehliswa kakhulu: Uma kuqhathaniswa nezinsizakalo zama-sandbox ezisingathwayo ezifana ne-E2B ne-Daytona, izindleko ze-Orchard Env ngokusetshenziswa zicishe zibe ngu-0.47×, kanti izindleko zama-Spot instance zingu-0.10× kuphela, okuwumehluko ocishe ube izikhathi ezingu-10.
Ikhono eliqinile lokujwayela ama-Harness ahlukahlukene: Ungqimba lwendawo olufanayo lusekela ama-harness amaningi afana ne-ReACT, ZeroClaw, OpenClaw ne-Codex. Ngaphansi kwe-ZeroClaw, i-Orchard-Claw ikhuphula i-pass@3 isuka ku-59.6% iye ku-73.9%.
Idatha yokuqeqesha esebenza kahle: I-Orchard-GUI ifinyelela ezingeni le-Agent yewebhu evulekile ehamba phambili isebenzisa imiboniso ye-distillation engu-400 kuphela nemisebenzi evulekile engu-2,200, okufakazela ukuthi idatha encane ehlanganiswe nendawo efanele ingaqeqesha imodeli enamakhono aphezulu.

Iwebhusayithi yephrojekthi：https://www.microsoft.com/en-us/research/blog/orchard-an-open-framework-for-scalable-agentic-ai/
Indawo yokugcina ye-GitHub：https://github.com/microsoft/Orchard
Indawo yamamodeli e-HuggingFace：https://huggingface.co/datasets/microsoft/Orchard
Iphepha lobuchwepheshe le-arXiv：https://arxiv.org/pdf/2605.15040

Ubunjiniyela besoftware obuzenzakalelayo: Ihlonza ama-bug ngokuzimele kumarepository ekhodi angempela, ibhale izivivinyo, ikhiqize ama-patch futhi iwaphase ekuqinisekisweni. Ilungele ukugcinwa kwamaphrojekthi avulekile nokubuyekezwa kwekhodi yangaphakathi ezinkampanini.
Ukuzulazula okuhlakaniphile kuwebhu: Isebenzisa ukuqonda okubonakalayo ukusebenzisa i-browser ngokuzenzakalelayo ukuze iqede imisebenzi yezizinda ezivulekile efana nokubhukha amathikithi, ukuthenga nokusesha ulwazi, futhi ingasetshenziswa njengenye indlela ye-RPA.
Umsizi womuntu siqu wokukhiqiza: Wenza ama-workflow ayinkimbinkimbi kuwo wonke amathuluzi afana ne-imeyili, ikhalenda namadokhumenti, njengokuhlela imihlangano ngokuzenzakalelayo, ukuhlela ibhokisi lokungenayo nokukhiqiza imibiko.
Ingqalasizinda yocwaningo lwe-Agent: Inikeza amaqembu ocwaningo ezemfundo nezimboni inkundla ephinde ikhiqizwe, enezindleko eziphansi, yokuqeqesha nokuhlola ama-Agent, isheshise ukwakhiwa kwe-ecosystem ye-Agentic AI evulekile.
Ukudluliswa kwamakhono e-Agent phakathi kwezizinda: Isebenzisa ungqimba lwendawo oluhlanganisiwe ukuhlola ukudluliswa kwamakhono nokufunda okuqoqekayo phakathi kokusebenza ngekhodi, amawebhusayithi nama-desktop.