Microsoft Unveils Orchard Framework for Cost-Effective AI Agent Training

Microsoft Research has released Orchard, an open-source framework designed to make training autonomous AI agents more cost-effective and easier. The tech giant's research division published Orchard on Monday, following its initial publication in March.

Orchard aims to address the persistent bottleneck faced by researchers in building state-of-the-art agentic systems, which often require proprietary infrastructure that most cannot access or reproduce. To achieve this, Orchard uses a lightweight Kubernetes environment called Orchard Env, providing reusable isolated components for training and evaluating agents at scale.

The framework includes three domain-specific training recipes: Orchard-SWE (software-engineering), Orchard-GUI (browser navigation), and Orchard-Claw (everyday productivity tasks). Microsoft released the results of models trained through each recipe, which showed they were competitive on established benchmarks.