Microsoft Orchard: Scalable AI Agent Training Framework

Microsoft Research has released Orchard, an open-source framework for training and evaluating AI agents. This new system helps researchers avoid rebuilding infrastructure for each project. Here's the thing: Orchard centers on Orchard Env, a lightweight, Kubernetes-native service. It supports data collection, reinforcement learning, and evaluation across software engineering, web navigation, and personal assistant tasks. The release includes three specific training recipes: Orchard-SWE, Orchard-GUI, and Orchard-Claw, along with their training data and evaluation methods. What's interesting is that Orchard-SWE, applied to software engineering tasks, improved performance on SWE-bench Verified from a 61.4% baseline to 73% with additional techniques. This system uses approximately three billion active parameters. Orchard-GUI, for browser tasks, recorded an average of 68.4% across three web-navigation benchmarks. The bottom line: This framework aims to make AI agent development more efficient and scalable for a variety of applications.