Quick answer
Orchard addresses a train-deploy mismatch: many open training stacks use a simplified agent loop even though the deployed agent runs inside a stateful, multi-process harness. Microsoft describes training rollouts in systems such as Codex, OpenClaw, and ZeroClaw by:
- launching each rollout in an Orchard Env container;
- running the real harness inside that environment;
- using a lightweight proxy to record the harness’s model calls;
- reconstructing those calls as samples for the training system.
This aligns data collection with the harness’s actual context and tool behavior. It does not make the model, harness, or tools interchangeable.
Keep four layers separate
| Layer | Examples | What must be pinned |
|---|---|---|
| Model | open-weight policy or teacher | checkpoint, tokenizer, inference settings |
| Harness | Codex, OpenClaw, ZeroClaw, ReACT | version, prompt, context policy, tool protocol |
| Environment | Orchard Env container | image, resources, network, credentials, lifecycle |
| Trainer | SFT or RL stack | code revision, sampler, reward, optimizer, seeds |
The Orchard repository currently says every sandbox includes codex, claude, pi, opencode, and hermes on PATH. The Microsoft Research introduction discusses training demonstrations across Codex, OpenClaw, and ZeroClaw. Do not assume every named harness is built into every image or release; verify the current repository and image.
A controlled comparison
To test whether real-harness training helps:
- define one frozen task and environment set;
- run the same model in the simplified training loop and target harness;
- compare tool schemas, context construction, retries, and stopping behavior;
- train on the target harness while holding the reward and task set fixed;
- evaluate both in the target harness on an untouched holdout;
- report compute, attempts, failures, and human-reviewed outcomes;
- test a second harness to measure transfer instead of assuming it.
Proxy capture can include prompts, tool arguments, outputs, repository content, or personal data. Apply data minimization, access control, retention, redaction, and deletion requirements before recording trajectories.
Deployment boundary
Harness-aligned training may reduce behavioral mismatch, but production still needs scoped tool permissions, explicit confirmation for sensitive actions, idempotency, rollback, observability, and human takeover. A successful rollout is not proof that the action was authorized.
Read the Orchard Env guide for isolation and the Claw guide for assistant-specific authority. The SWE guide covers code-agent evaluation.
Frequently asked questions
Why train an agent inside its deployment harness?
A harness controls multi-turn context, tools, retries, and external connections. Training in a simplified imitation can create a mismatch between the trajectories learned during training and the system used during evaluation or deployment.
How does Orchard collect training data from a real harness?
Microsoft describes a lightweight proxy that records the harness’s own model calls while Orchard Env runs each rollout in a separate container, allowing those calls to be reconstructed as training samples.
Does Orchard make Codex or OpenClaw safe to deploy?
No. Harness-aligned training does not replace model evaluation, tool permissions, data controls, user confirmation, monitoring, or deployment approval.
Official sources
Source check: August 4, 2026. Verify current harness inclusion, licenses, model-call capture, trainer compatibility, image contents, credentials, and retention before collecting trajectories.