Microsoft Research released Orchard, an open-source toolkit for training and testing AI agents, together with three training recipes for software engineering, web browsing and personal-assistant tasks. Its core piece runs thousands of isolated sandboxes on Kubernetes and can train an agent inside the same tool it will later run in, such as Codex or OpenClaw. Training data and evaluation methods are published as well, and results come from small open-weight models rather than frontier systems.
What changed
Building competitive AI agents generally required in-house sandboxes, closed training pipelines and private datasets that outside researchers could not reproduce.
What it unlocks
Training and evaluating agents inside the same real command-line and browser tools they will be deployed with, using shared, reusable sandbox infrastructure rather than a simplified stand-in.
- Orchard-SWE: 69.7% on SWE-bench Verified, 73.0% with value-model reranking
- about 3 billion active parameters, versus frontier systems more than 10x larger
- Orchard-GUI (4B): 74.1% WebVoyager, 67.0% Online-Mind2Web, 64.0% DeepShop, 68.4% average
- Orchard-Claw: 59.6% on Claw-Eval, 73.9% with ZeroClaw; Codex harness 18.6% untrained to 51.5% trained
- 107,000 distilled agent interactions; Orchard-GUI trained on 400 demonstrations plus 2,200 tasks
What you need to act on it
- a Kubernetes cluster to run the environment service
- GPU capacity for fine-tuning and reinforcement learning
- microsoft.com2026-08-03