Orchard: An open framework for scalable agentic AI

Microsoft Research has unveiled Orchard, an open-source framework aimed at advancing research in autonomous, agentic AI systems. Central to Orchard is Orchard Env, a reusable, Kubernetes-based environment service that enables researchers to train and evaluate AI agents across diverse task domains, including software engineering, web navigation, and personal assistant applications.

Unlike traditional frameworks, Orchard Env is designed to be a standalone, flexible service that supports different agent architectures and task types without requiring modifications. It facilitates training agents directly within real deployment harnesses such as Codex, OpenClaw, and ZeroClaw, allowing for end-to-end training that closely mirrors real-world deployment conditions.

The project includes domain-specific training recipes—Orchard-SWE for software engineering, Orchard-GUI for web navigation, and Orchard-Claw for personal assistant tasks—along with the associated training data and evaluation methods. For example, Orchard-SWE, trained on 107,000 interactions from models like MiniMax-M2.5 and Qwen3.5-397B, achieved a state-of-the-art 69.7% accuracy on the SWE-bench Verified benchmark with only about 3 billion parameters, approaching larger models.

Similarly, Orchard-GUI, a vision-language model trained on limited supervision, demonstrated strong performance on web navigation benchmarks, achieving an average success rate of 68.4%. Orchard-Claw, focused on everyday productivity tasks, successfully completed nearly 60% of tasks on the Claw-Eval benchmark, with performance improving significantly when paired with stronger systems like ZeroClaw.

A key innovation of Orchard is its environment layer, which manages isolated components for data collection, reinforcement learning, and evaluation, supporting multiple stages of agent development without rebuilding infrastructure. This approach reduces the barriers to research and accelerates experimentation, enabling the community to build more capable open agents efficiently.

Looking forward, Orchard emphasizes the importance of reusing training experiences to enable cumulative learning, where agents can inherit and extend previous knowledge. The release of the full Orchard stack aims to democratize access to scalable agentic AI research tools, fostering innovation and collaboration across the community.