AIBRIEF
Microsoft Research releases an open framework for training AI agents that plan and act across multi step tasks — and a ~3B parameter model that nearly matches much larger systems on a real coding agent benchmark.
Microsoft Research has released Orchard, an open framework for training and evaluating agentic AI: systems that plan and act across multi-step tasks like fixing code, browsing the web, or handling email and calendar work, rather than just answering questions. The release is described in a Microsoft Research blog post.
The framework's centerpiece is Orchard Env, a Kubernetes-based service that supplies reusable, isolated environments for training, reinforcement-learning rollouts, and evaluation. The same service supports software-engineering, web-browsing, and personal-assistant agents without modification, and the project ships training recipes (Orchard-SWE, Orchard-GUI, Orchard-Claw), training data, and evaluation methods so outside teams can reuse the pipeline.
Microsoft Research is also releasing a roughly 3-billion-active-parameter open-weight model that, per the blog post, hits 69.7% on SWE-bench Verified, a benchmark measuring real coding-agent behavior, and 73.0% with value-model reranking. The source frames that result as approaching frontier systems built on more than 10 times larger models. The same Orchard infrastructure can train agents inside third-party deployment harnesses, including Codex, OpenClaw, and ZeroClaw.
The benchmark numbers are self-reported by Microsoft's researchers, and SWE-bench Verified measures coding-agent behavior, not general real-world usefulness. Independent reproduction and real-deployment performance remain unverified.