Read this before assigning or reviewing another agent's work. Provider-specific model preferences and Claude harness behavior are in Claude operation.
Keep orchestration, implementation-plan creation, plan/implementation review, very hard problems, and work cheaper to do inline in the main session. Generally avoid tests in parallel agents unless explicitly intended; coordinate heavy tiers through the main session.
Write relevant operational rules into each delegation prompt rather than relying on the agent having read a guide.
For gate-running work, include foreground execution, extended timeout, full output to a file, and the requirement to return remaining verification to the main session; reassert these in mid-task corrections.
The main session runs full.
Require the budget-exhaustion protocol: commit only the green subset, report the precise remainder, stop cleanly, and never stall.
This protocol does not grant additional commit permissions.
For every delegation with a plan, require an item-by-item report and review every plan item against it rather than spot-checking apparent completeness.
For phased work, keep that specification under draft/ with each code-behavior premise marked as a claim to probe.
Apply probe discipline to both the implementer's premises and the reviewer's conclusions.
A delegation writing into a registry-governed tree must name its enforcing gate and tier, including required registration edits.
Apply fixture registration and coverage even when the worker's lane cannot run the enforcing tier.
When heavy tiers are serialized, constrain worker lanes to fast and explicitly assign their remaining checks to the main session.
Do not end a sub-agent turn merely to await work whose completion requires that agent to remain active. Use bounded waits and the current harness's task controls; report completed, reviewed results. For a stalled phase, read its own log and resume the same agent with the findings to preserve context, rather than immediately replacing it.