Microsoft Open-Sources Orchard to Train Agents in Real Harnesses

Microsoft Research open-sourced Orchard, a framework that trains AI agents inside the real harnesses they'll actually run in, not a simplified stand-in.

Most agent reinforcement learning happens in a simplified environment. The mismatch shows up once the agent deploys into a real tool like Codex or OpenClaw. Orchard trains the agent end-to-end inside the harness itself, closing that gap before deployment.

The core piece is Orchard Env, a Kubernetes-based environment service that hands out reusable, isolated components for data collection, RL rollouts, and evaluation. It's built to work across domains without per-domain rework.

The mechanism is a lightweight proxy that sits inside the harness and records its own model calls as training data. Each rollout runs in its own container. That's what lets an agent train directly inside Codex or OpenClaw, instead of a mock version of either.

Microsoft is shipping three training recipes with the framework, along with training data and evaluation methods, per its research blog.

Harness fidelity over more RL compute

The bet here is that the gap between training environment and deployment environment, not a shortage of RL compute, has been capping agent quality. If that holds, agents trained on simplified stand-ins keep losing to harness-trained ones once both hit production, benchmarks aside.

Worth watching whether other labs open their harnesses to the same proxy trick, or whether this stays specific to Codex and OpenClaw.