Microsoft's Orchard framework trains agentic AI with 3B params near frontier

Microsoft Research has introduced Orchard, an open-source framework designed to scale agentic AI research. At its core is Orchard Env, a lightweight Kubernetes-based environment service that provides reusable, isolated infrastructure for training and evaluating autonomous agents across diverse domains—from fixing bugs in codebases to navigating the web and managing personal-assistant workflows. Unlike proprietary systems locked behind custom sandboxes and closed pipelines, Orchard lets researchers collect training data, run reinforcement learning rollouts, and execute evaluations without rebuilding infrastructure for each new task or agent system.

The framework's standout capability is its support for training agents directly inside production deployment harnesses such as Codex, OpenClaw, and ZeroClaw. By using a lightweight proxy to record model calls during real rollouts, Orchard eliminates the mismatch between simplified training environments and real-world deployment. Demonstrating its efficiency, the Orchard-SWE model reaches 69.7% on SWE-bench Verified—73.0% with value-model reranking—while using only about 3 billion active parameters, approaching frontier systems that are more than 10 times larger. Alongside recipes for Orchard-SWE, Orchard-GUI, and Orchard-Claw, the project releases training data and evaluation methods, aiming to make reproducible, state-of-the-art agentic AI research accessible to the broader community.