Microsoft Research has released Orchard, an open-source framework for building, training, and evaluating autonomous AI agents that the company says lets relatively small open-weight models approach the performance of frontier systems more than ten times their size. The project, detailed in a blog post on August 3, 2026, is the company's most ambitious bid yet to give the broader research community the kind of infrastructure that has, until now, largely remained locked inside proprietary labs.
The release arrives as the AI industry shifts from static question-answering toward agents that can plan, reason, and act across multistep environments. For more on how the field is moving toward autonomous systems, see our latest AI industry coverage.
What Orchard actually is
At the center of the project is Orchard Env, a lightweight, Kubernetes-based environment service that provides reusable, isolated components for running agents at scale. According to Microsoft, the same service can support software-engineering agents, web-browsing agents, and personal-assistant agents without modification. It handles everything from collecting training data to reinforcement-learning rollouts and evaluation.
The key architectural decision is that the runtime environment is treated as a standalone, reusable service rather than infrastructure embedded inside a specific training framework. Because Orchard Env sits on Kubernetes, it can create, manage, and remove thousands of isolated containers in parallel. Teams can introduce new benchmarks, agent systems, or training algorithms without rebuilding the underlying infrastructure from scratch.
Training inside real harnesses
One of Orchard's distinguishing features is its ability to train agents directly inside the deployment harnesses they will actually use in production. Today's most capable agents rarely run as a bare model. They operate through sophisticated harnesses such as Codex, OpenClaw, and ZeroClaw that manage multi-turn reasoning, tool use, and connections to external systems.
Open training tools usually cannot handle these stateful, multi-process harnesses, forcing researchers to train on a simplified stand-in and then deploy in the real setting, which creates a mismatch between training and deployment. Orchard closes this gap: a lightweight proxy records the harness's own model calls as training data while each rollout runs in its own container, allowing an agent to be trained end-to-end directly in the harness it will use in production.
Three domain-specific recipes
To demonstrate the approach, Microsoft released three training workflows.
Orchard-SWE: software engineering
Orchard-SWE targets one of the most demanding settings for autonomous agents: multi-step reasoning over real codebases. It was built using the Mini-SWE-Agent framework and evaluated on SWE-bench Verified, a benchmark that tests a model's ability to navigate, diagnose, and repair real-world codebases.
To train the system, Microsoft distilled 107,000 agent interactions from two advanced open-weight models, MiniMax-M2.5 and Qwen3.5-397B, covering a broad range of GitHub issues. Using a technique called credit-assignment supervised fine-tuning, the system learns from the productive portions of partial attempts rather than discarding them entirely.
The results move the model from a 61.4% baseline on SWE-bench Verified to 69.1% with reinforcement learning, and 69.7% with dense-reward techniques. With value-model reranking, the figure reaches 73.0% — a new state of the art among open-source models of comparable size, using only about 3 billion active parameters.
Orchard-GUI: web navigation
Orchard-GUI trains a 4-billion-parameter vision-language model as a browser agent. Despite using a relatively small amount of supervision — 400 distilled demonstrations combined with 2,200 open-ended training tasks — the model achieved 74.1% on WebVoyager, 67.0% on Online-Mind2Web, and 64.0% on DeepShop, for an average of 68.4%. Microsoft says these results place it among the strongest open-source web agents to date.
Orchard-Claw: personal assistant tasks
Orchard-Claw se concentre sur les tâches de productivité quotidiennes telles que la lecture et la rédaction d'e-mails, la gestion de calendriers et la recherche d'informations. Entraîné sur seulement 200 tâches synthétiques et évalué sur le benchmark Claw-Eval, il a réalisé 59,6 % des tâches assignées jusqu'à trois tentatives, passant à 73,9 % lorsqu'il est associé au système d'agent ZeroClaw plus puissant.
Pourquoi les résultats des petits modèles sont importants
Les chiffres de référence sont remarquables car ils proviennent de modèles comportant environ 3 à 4 milliards de paramètres actifs, soit bien plus petits que les systèmes frontières d'OpenAI, Anthropic et Google qui dominent les classements. L'argument de Microsoft est qu'une infrastructure et un environnement de formation adéquats peuvent combler une grande partie du fossé, en rendant les systèmes agents performants accessibles aux chercheurs et aux organisations qui n'ont pas les moyens de former ou d'exécuter les plus grands modèles propriétaires.
Parallèlement aux modèles et aux flux de travail, Microsoft a publié les données de formation et les méthodes d'évaluation utilisées pour les créer, dans le but déclaré d'aider la communauté de recherche au sens large à créer et à étudier des systèmes agents ouverts. Les travaux ont été dirigés par Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng et Jianfeng Gao de Microsoft Research.
Gardez une longueur d'avance sur l'IA
La version Orchard de Microsoft signale que l'infrastructure derrière l'IA agentique de frontière commence à se démocratiser. En savoir plus dernières actualités sur l'IA et Lire plus d'actualités sur l'IA →
