Microsoft released Orchard to standardize the training of autonomous agents

Microsoft introduced Orchard, an open-source framework designed to help researchers build and evaluate autonomous agents. The system provides a Kubernetes-based environment called Orchard Env that manages isolated components for training and testing. This approach enables developers to reuse infrastructure for different task types, such as coding or web navigation, without building new tools for every project.

The framework includes three specific training recipes: Orchard-SWE for software engineering, Orchard-GUI for web tasks, and Orchard-Claw for productivity. By allowing agents to train directly within deployment harnesses like Codex or OpenClaw, the system reduces the performance mismatch that often occurs between training and real-world execution. Orchard-SWE achieved a 73% success rate on the SWE-bench Verified benchmark using a 3-billion-parameter model.

Researchers tested these agents on specific benchmarks, showing that smaller models can compete with larger systems when trained effectively. The team emphasizes that reusing training data and trajectories can accumulate agent experience over time. However, the system currently requires significant computational resources to run parallel containers, and its success is limited to the specific tasks and benchmarks provided in the release.