Orchard: An Open-Source Framework for Scalable Agentic AI

Microsoft has released Orchard, an open-source framework for scalable and cost-efficient agentic AI research. Its core component, Orchard Env, is a lightweight Kubernetes-based environment service that provides isolated execution environments at scale for agent data collection, reinforcement learning rollouts, and evaluation.

Orchard Env is designed to be agnostic to specific agents or task types. It supports software engineering, web navigation, and personal assistant agents on the same infrastructure, and enables training directly inside real deployment harnesses such as Codex, OpenClaw, and ZeroClaw.

The project also released three domain-specific training recipes along with related data and evaluation methods.

• Orchard-SWE: Using over 107,000 trajectories, it achieved 67.5% on SWE-Bench. A model with approximately 3 billion active parameters achieved 69.7% on SWE-bench Verified, and 73.0% with value-model reranking applied.
• Orchard-GUI: Scored 74.1 on WebVoyager, 67.0 on Online-Mind2Web, and 64.0 on DeepShop.
• Orchard-Claw: Achieved 31.7 pass and pass@3 59.6, a 19.9-point improvement over the base model.

Orchard focuses on opening up agentic AI research infrastructure that has relied on proprietary sandboxes and closed data pipelines, enabling researchers to reuse environment, data, and evaluation workflows across multiple task domains.