Microsoft Research has released the open-source Orchard framework for training AI agents

The research division Microsoft Research has released an open-source framework called Orchard, which aims to simplify and reduce the cost of training and evaluating autonomous AI agents. The release follows an earlier publication on this topic from March. According to Microsoft, the framework eliminates the need for researchers to build their own sandbox infrastructure, data pipelines and evaluation systems for individual models and use cases.

At the core of the framework is Orchard Env, which Microsoft describes as a lightweight Kubernetes environment with reusable isolated components. These can be used to collect training data, run reinforcement learning and evaluate agents at scale without having to rebuild the underlying architecture.

Alongside the framework, Microsoft released three domain-specific training workflows, including training data and evaluation methods: Orchard-SWE, focused on agents for software engineering; Orchard-GUI, for browser and GUI navigation; and Orchard-Claw, for everyday productivity tasks. According to Microsoft, models trained using these workflows achieved competitive results on established benchmarks.

Microsoft stated in a blog post that the agentic AI research community faces a persistent problem, as building state-of-the-art agent systems typically requires proprietary infrastructure unavailable to most researchers and practitioners. According to the company, by making the infrastructure open, lightweight and reusable, Orchard reduces the cost of agentic AI research because teams no longer have to build their own isolated environments from scratch or rely on proprietary cloud services. The company also hinted at a future direction for development – reusing training trajectories as lasting assets, for example by distilling them into reusable value models, instead of discarding them after a training run ends.