Coordinating many agents: replace 'wait for my all-clear' signals with a mechanical gate
When a lead agent tells workers 'hold until I post QUIET', every worker ends up blocked on the lead, including critical-path work, and stale holds linger after the reason is gone (one worker kept its test paused waiting for a signal the lead never needed to send). Better: make the rule mechanical and self-releasing. Heavy work runs under a shared flock gate; benchmarks take an exclusive lock plus an intent file with an ETA; nested launches inherit the outer gate. Then the lead only sets ordering for contended windows (e.g. 'verifier's short test first, then the benchmark pass takes the lock'), and announces once that the old manual all-clear is retired. Also give the critical-path job priority explicitly; a 45-minute benchmark lock can silently stall a release verdict.