Microsoft Research releases Orchard, an open framework for training and evaluating agentic AI systems

Microsoft Research has released Orchard, an open-source framework aimed at simplifying the development and evaluation of AI agents across diverse task domains. The framework is designed to reduce complexity by providing reusable infrastructure, allowing researchers to focus on model development rather than rebuilding environments from scratch.

At the core of Orchard is Orchard Env, a reusable environment service built on Kubernetes that supports training and evaluating agents across domains such as software engineering, web navigation, and personal assistance. The infrastructure is designed to work with existing agent systems and harnesses, including Codex, OpenClaw, and ZeroClaw, enabling end-to-end training within real deployment environments.

The framework includes three domain-specific workflows—Orchard-SWE, Orchard-GUI, and Orchard-Claw—demonstrating its versatility. For example, Orchard-SWE, a workflow for software engineering agents, achieves 69.7% on the SWE-bench Verified benchmark using a model with approximately 3 billion active parameters. With value-model reranking, performance improves to 73.0%, approaching the results of frontier systems that use models more than ten times larger.

Orchard-SWE is trained using 107,000 agent interactions distilled from two advanced open-weight models, MiniMax-M2.5 and Qwen3.5-397B, covering a broad range of GitHub issues. The training process employs credit-assignment supervised fine-tuning to learn from partial successes, expanding the amount of useful training data. Reinforcement learning is then applied using techniques such as Balanced Adaptive Rollout, on-policy distillation, and process rewards to provide denser feedback signals.

The project also releases training data and evaluation methods alongside the framework, providing resources for the broader research community to build and study open agentic systems. This release underscores Microsoft Research’s focus on advancing open, reproducible research in autonomous AI systems.