Orchard: An open framework for scalable agentic AI

Microsoft Research has released Orchard, an open-source framework for scalable and cost-effective agentic AI research. At its core is Orchard Env, a Kubernetes-based environment service providing reusable isolated components for training and evaluating agents across task domains. The framework supports software-engineering, web-navigation, and personal-assistant agents, and can train them directly inside real deployment harnesses such as Codex, OpenClaw, and ZeroClaw. The project introduces three training recipes: Orchard-SWE, Orchard-GUI, and Orchard-Claw. For software engineering, Orchard-SWE uses 107,000 distilled agent interactions from MiniMax-M2.5 and Qwen3.5-397B, with credit-assignment supervised fine-tuning and reinforcement learning, reaching 69.7% on SWE-bench Verified (73.0% with value-model reranking) using only about 3 billion active parameters. For web navigation, Orchard-GUI trains a 4-billion-parameter vision-language model with limited supervision (400 demonstrations and 2,200 tasks), achieving an average of 68.4% on WebVoyager, Online-Mind2Web, and DeepShop. For personal assistance, Orchard-Claw, trained on 200 synthetic tasks across multiple harnesses, achieves 59.6% on Claw-Eval (73.9% with ZeroClaw). The framework aims to lower research costs by providing open infrastructure, and the authors highlight data efficiency and cumulative agent learning as future directions. All training data and evaluation methods are released to the community.