Microsoft published Orchard, an open-source framework for training and testing software agents, under an MIT license. It centres on a Kubernetes-based sandbox service with a Python client that creates large numbers of isolated containers on demand, plus three training recipes for coding, browser and assistant tasks and two public trajectory datasets. Using it requires standing up a cluster, since Microsoft provides code and setup scripts rather than a hosted service.
What changed
Teams training agents on isolated test environments generally rebuilt the sandbox layer inside each training stack or paid for managed sandbox services.
What it unlocks
Running thousands of isolated containers on an own Kubernetes cluster to train and evaluate agents, and reusing the same environment across different agent command-line tools without rebuilding images.
- 107,185 multi-turn software-engineering rollouts across 2,788 repositories
- 3,070 judge-verified browser-navigation rollouts
- 73.0% on SWE-bench Verified (Orchard-SWE)
- 0.28 s average command-execution latency vs 0.747 s E2B and 2.046 s Modal
- 1,000 sandboxes launched in parallel, 100% success, 26 s end-to-end
- $673 on spot for 128 sandboxes x 240 h vs $7,078-$10,305 managed services
What you need to act on it
- a Kubernetes cluster (four scripts provision one on Azure AKS in about 20 minutes)
- own training hardware and stack for the reinforcement-learning recipes
- github.com2026-08-03