Orchard: An open framework for scalable agentic AI

Details

Microsoft Research's Orchard framework addresses a reproducibility bottleneck in agentic AI: most state-of-the-art agentic systems require proprietary infrastructure, closed training pipelines, and datasets inaccessible to most researchers. Orchard centers on Orchard Env, a lightweight Kubernetes environment that provides reusable isolated components — from data collection to reinforcement learning rollouts and evaluation — that work across task domains without modification. Three released domain recipes demonstrate the scope: Orchard-SWE (software engineering), Orchard-GUI (web navigation), and Orchard-Claw (personal assistant). Orchard-SWE achieves 69.7% on SWE-bench Verified — rising to 73.0% with value-model reranking — using approximately 3 billion active parameters, compared to frontier systems using models more than 10 times larger. This efficiency result is the most concrete technical claim in the release. The framework runs directly inside deployment harnesses including Codex, OpenClaw, and ZeroClaw, letting researchers reuse environments, data pipelines, and evaluation workflows without rebuilding infrastructure per task. Training data and evaluation methods are released alongside model weights — addressing the full reproducibility chain rather than just checkpoints.