Microsoft Research has open-sourced Orchard, a framework built to train and evaluate AI agents across very different jobs without rebuilding the underlying infrastructure. The centerpiece, Orchard Env, is a Kubernetes-based service that spins up thousands of isolated environments in parallel, so the same setup can serve software-engineering agents, browser agents, and personal-assistant agents.
The project also tackles a mismatch that has frustrated agent researchers: models are usually trained on simplified stand-ins and then dropped into real tools. Orchard records a harness’s own model calls during rollouts, which lets teams train agents directly inside Codex, OpenClaw, or ZeroClaw, the environments they will actually run in.
Three training recipes ship with the release. Orchard-SWE reached 69.7 percent on SWE-bench Verified, or 73 percent with value-model reranking, using only about 3 billion active parameters, a result that approaches frontier systems more than ten times larger. It was built by distilling 107,000 agent interactions from MiniMax-M2.5 and Qwen3.5-397B, then adding reinforcement learning with a process reward model that rewards sound problem-solving rather than only final answers. Orchard-GUI, a 4-billion-parameter browser agent trained on just 400 demonstrations, averaged 68.4 percent across WebVoyager, Online-Mind2Web, and DeepShop.
Microsoft is releasing the training data and evaluation methods alongside the code, in the hope that open agent research can stop depending on proprietary sandboxes and closed pipelines.
