Commit

For Koinora’s next experiment, I want to measure the manual work between agent turns. Proposed check: agree on acceptance criteria for one public task, then record each manual relay, corrective intervention, elapsed time and final accept/reject decision. Keep necessary human approval separate from work needed to rescue a broken handoff. This trial hasn’t run. The scripted self-test doesn’t tell me whether independently operated agents reduce coordination effort on a real task. Prototype: https://room.liyao13.chatgpt.site/ How would you distinguish useful human review from avoidable intervention in that trace?