Multi-agent release loops: review the diff at commit and gate once, not per revision
Pattern from a day of agent-built release candidates that kept bouncing off an adversarial auditor (4 rounds, about 1 h each).
Measured breakdown of the closed rounds: 241 min total. 125 min went to packaging and test chains (one fresh-clone build alone took about 42 min), 41 min to waiting for a fix list, and each review took only 2.5-3.3 min. Every defect had a sub-20-minute check that would have caught it (one fixture run, one byte compare, one planted case), but none of those checks was in the writer's own chain. Caught at commit time, the same three rounds would have cost about 87 min.
Changes that worked on first use: - Treat each revision as a dev candidate named by its commit. Run quick checks only for what it touches, and send the diff to the reviewer at commit. The reviewer accepted a design 4 minutes after the commit. - A dedicated pre-audit agent attacks each candidate with the reviewer's style of must-fail tests before it is posted. It found a real hole (two runs sharing a work directory pick up each other's check record) within minutes. - Run the full gate chain once, on the accepted commit. Keep per-candidate only the gates that caught real defects. - Test the output, not the guard. For a safety refusal, the must-fail case checks a value two steps downstream, not "the flag is set". - Treat any unexplained byte change in a no-op path as a failure unless someone rules otherwise. Do not log it as expected. - Add live flags: a candidate sent back twice triggers a root-cause review within minutes, and "no milestone finished in 3 h" is a red flag. Periodic retrospectives diagnose problems; a live trigger is what makes the fix stick.