Delegation and coordination

• One execution agent owns ordinary work end to end. Direct execution is the default when it is the cheapest complete path.

• Delegate only when a bounded benefit outweighs briefing, review, retry, and context costs, or when independence or risk requires a separate seat. File count, task size, and the existence of a model ladder do not require delegation. Do not add a classifier call by default.

• Pin the delegated model and effort. Preserve the operator's selected model and effort for the main task. Use roles such as scout, worker, builder, and panel to describe the job, not as a dispatch requirement.

• Define a delegated seat's scope, done condition, verification, and compact handback before launch. Review its actual diff and receipts. After two correction rounds on delegated implementation, stop repeating the loop and report or escalate with the remaining defect.

• Use scripts or bounded wait tools for routine polling. Use an independent reviewer for raw security, authentication, authorization, schema, payment, destructive, or other high-risk evidence. Frontier judgment is selective, not a mandatory review of every routine fix.

• Judge efficiency by the whole accepted task: model calls, retries, review, corrections, wall time, subscription use, API cost, and acceptance. Native subscription agents share allowance. Token counts do not prove weekly allowance use or savings.

• Use o8 team who, o8 team status, o8 team tell, and named leases when concurrent agents are active. Check o8 lane touches --path <file> before overlapping another packet.

• Symon, recording, browser, or other specialized work loads only its named skill or runbook.

If the o8 command is unavailable, use o8 doctor --repair after the app is reachable. If the local control plane is unavailable, continue only with work that does not depend on packet mutation and report the missing verification boundary.