Microsoft published Orchard, an openly licensed framework for building and testing AI agents, covering a Kubernetes-based sandbox service, a Python client, and reinforcement-learning training code. It also released trajectory datasets and results for three agent recipes covering software repair, browser navigation, and assistant workflows. Using it requires standing up a cluster, for which deployment scripts targeting Azure are included.
What changed
Teams training or testing coding agents relied on managed sandbox services or ad-hoc Docker setups rebuilt per project.
What it unlocks
Running thousands of isolated agent sandboxes on a self-hosted Kubernetes cluster, with several agent command-line tools preinstalled, and reusing the same setup for both training and evaluation.
- 0.28 s average command latency
- $673 on spot vs $7,078+ managed
- 73.0% SWE-bench Verified
- 107,185 SWE trajectories released
What you need to act on it
- a Kubernetes cluster (scripts provided for Azure AKS)
- engineering capacity to deploy and operate the service
- GPU training stack for the reinforcement-learning recipes
Sources