Microsoft Research Releases Orchard, an Open Framework for Agentic AI

Microsoft Research has released Orchard, an open-source framework intended to make research on scalable agentic AI more reusable across task domains.

A shared foundation for agent research

According to the Microsoft Research blog, Orchard is built around Orchard Env, a reusable environment service for training and evaluating agents. The framework is designed to support work on different kinds of agents, including systems for software engineering, graphical web navigation and personal-assistant workflows.

Microsoft says the same infrastructure can be used to train agents within deployment harnesses including Codex, OpenClaw and ZeroClaw. The stated goal is to let researchers reuse environments, datasets and evaluation methods rather than rebuilding task-specific foundations for each experiment.

The accompanying Orchard paper on arXiv describes the project as an open-source framework for “scalable agentic modeling.” It identifies three implementation recipes: Orchard-SWE for software-engineering tasks, Orchard-GUI for graphical-interface tasks, and Orchard-Claw for personal-assistant agents.

Focus on smaller open-weight models

Microsoft Research positions Orchard as evidence that comparatively small open-weight models can be effective on difficult agent tasks when paired with suitable training environments and methods.

The company reports that Orchard-SWE achieved 69.7% on SWE-bench Verified, a benchmark based on resolving issues in real software repositories. With value-model reranking, Microsoft reports a score of 73.0%. The system uses about 3 billion active parameters, according to the blog post.

Microsoft says that result approaches the performance of frontier systems that use models more than ten times larger. The comparison reflects the company’s own characterization and does not by itself establish equivalence across model capabilities, costs or deployment conditions.

Code, data and environments

The project release includes the environment foundation and links to accompanying trajectory datasets, according to Microsoft’s GitHub repository. Microsoft Research also says it is releasing training data and evaluation approaches alongside the models and workflows.

That combination addresses a practical challenge in agent research: agents must do more than generate text. They need environments where they can take actions, observe outcomes and be assessed on multistep work. Microsoft cites examples such as fixing bugs in complex codebases, navigating websites and managing workflows involving calendars and email.

By releasing a common framework rather than only individual models, Microsoft Research is aiming to give researchers a basis for comparing approaches across distinct forms of agent behavior. Orchard’s usefulness will depend on how broadly its environments and released resources are adopted, as well as whether its reported results can be independently reproduced on the cited benchmarks.

Sources

Microsoft Research, “Orchard: An open framework for scalable agentic AI”; the Orchard paper on arXiv; and Microsoft’s Orchard GitHub repository.