Quick answer
Orchard-Claw is the personal-assistant recipe in Microsoft Orchard. Its task domain includes reading and drafting email, managing calendars, searching for information, and coordinating across tools.
Microsoft reports training on 200 synthetic tasks, with 59.6% task completion on Claw-Eval when the agent receives up to three attempts and 73.9% when paired with ZeroClaw. The Microsoft post also reports that, under a Codex harness, the measured success rate rose from 18.6% for the untrained model to 51.5% after Orchard training.
These are research results for specified models, harnesses, attempts, and evaluation rules. They do not transfer action authority or prove reliability on a real user’s inbox or calendar.
Why the harness matters
The same model can behave differently under ReACT, OpenClaw, ZeroClaw, Codex, or another agent system because the harness controls context, tool exposure, retries, and multi-turn flow. Orchard’s environment abstraction allows a recipe to train across multiple harnesses instead of assuming one simplified loop.
Use the real-harness training guide to separate the model, harness, tools, and environment in your experiment.
Personal-assistant safety gates
Build the evaluation around authority, not only task completion:
| Gate | Required decision |
|---|---|
| Data access | Which accounts, folders, calendars, and fields may be read? |
| Action scope | May the agent draft, send, reschedule, invite, or delete? |
| Destination | Which recipients and external systems are approved? |
| Confirmation | Which actions require a user preview and explicit approval? |
| Idempotency | How are duplicate messages or calendar changes prevented? |
| Audit and recovery | Can an operator reconstruct, revoke, or undo the action? |
Use synthetic or isolated accounts first. Include ambiguous names, stale threads, conflicting calendars, prompt injection, sensitive attachments, and revoked permissions. Score correct abstention and escalation, not only completion.
Decide whether the recipe transfers
Reproduce a small task set with the same harness before changing the model. Then evaluate a different harness while holding the model and tasks fixed. This reveals whether gains come from training, the harness, or their interaction.
For shared sandbox design, read the Orchard Env guide. For the other domain recipes, use the Orchard overview.
Frequently asked questions
What is Orchard-Claw?
Orchard-Claw is Microsoft’s personal-assistant training recipe for workflows such as email, calendars, search, and coordination across tools, built on the shared Orchard environment.
What data and result does Microsoft report for Orchard-Claw?
Microsoft reports training on 200 synthetic tasks, 59.6% task completion on Claw-Eval with up to three attempts, and 73.9% when paired with the ZeroClaw agent system. The figures are setup-specific research results.
Can Orchard-Claw decide which user actions are authorized?
No. Training and benchmark success do not define permission to read data, send messages, change calendars, disclose information, spend money, or perform irreversible actions.
Official sources
Source check: August 4, 2026. Verify the model, harness, task generation, attempt budget, benchmark revision, tool permissions, and approval policy before use.