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

The Microsoft Research division released Orchard, an open-source framework for scalable training and evaluation of agentic AI systems. At its core is Orchard Env – a lightweight environment built on Kubernetes that provides reusable isolated components for running and building agents at scale, from training data collection through reinforcement learning rollouts to evaluation. According to Microsoft Research, unlike many existing frameworks, Orchard Env is designed to support different types of agents and tasks without modification – software-engineering agents, web-browsing agents and personal-assistant agents. The Kubernetes foundation enables thousands of isolated components to be created, managed and destroyed in parallel.

According to the company, the framework also addresses the problem of training within sophisticated harnesses such as Claude Code, Codex or OpenClaw, which manage multi-turn reasoning, tool use and connections to external systems. A lightweight proxy tool records model calls directly from the harness as training data, with each rollout running in its own container – allowing the agent to be trained end-to-end directly in the harness in which it will be deployed (OpenClaw, Codex, ZeroClaw and others), instead of in a simplified substitute.

The release includes three domain-specific training recipes – Orchard-SWE, Orchard-GUI and Orchard-Claw – along with the training data and evaluation methods used to create them. Orchard-SWE is built on the Mini-SWE-Agent framework and evaluated on the SWE-bench Verified benchmark. According to the company, 107 000 agent interactions were distilled for training from two open models (MiniMax-M2.5 and Qwen3.5-397B) across a broad sample of GitHub Issues, using supervised fine-tuning with credit assignment that also makes use of partially successful agent attempts. This is followed by reinforcement learning using Balanced Adaptive Rollout, supplemented with denser reward techniques (on-policy distillation and process reward model), and finally reranking candidate solutions using a 4-billion-parameter value model trained on previous rollouts.

According to Microsoft Research, this approach increased the Orchard-SWE score on SWE-bench Verified from a baseline of 61.4 % to 69.1 % (Balanced Adaptive Rollout), 69.7 % (with denser reward techniques) and 73 % after reranking with the value model – which the company describes as a new best result among open-source models of comparable size (approximately 3 billion active parameters), approaching the results of frontier systems more than ten times larger. Orchard-GUI trains a 4-billion-parameter vision-language model as a browser agent with a smaller amount of supervision (400 distilled examples). Details on Orchard-GUI and Orchard-Claw can be found in the source article.