Microsoft Launches Open-Source AI Agent Framework Orchard, Rivaling Systems 30x Its Size

Microsoft launches Orchard, an open-source AI agent framework that punches far above its weight, achieving benchmark scores rivaling systems up to 30 times larger while offering 107K+ training datasets and a Kubernetes-native sandbox for scalable agent development across software engineering, web navigation, and computer use.

• Microsoft has released Orchard, an open-source agentic modeling framework built on a Kubernetes-native sandbox service called Orchard Env, which enables scalable, isolated container-based environments for training and evaluating AI agents across software engineering, browser navigation, and computer use domains.
• Three research recipes — Orchard-SWE, Orchard-GUI, and Orchard-Claw — demonstrate strong benchmark results, with Orchard-SWE achieving 73.0% on SWE-bench Verified and Orchard-GUI reaching 68.4% average across live-web benchmarks, both approaching or matching systems 10–30 times larger in parameter count.
• The framework ships with 107K+ open-source trajectory datasets, a Python SDK, and a roadmap featuring stateful sandbox snapshotting and branching to enable more precise credit assignment in reinforcement learning, with all components designed to be portable across trainers, harnesses, and task domains.