Audited 1,228 human interventions in our AI agent setup. 91% weren't decisions, so we stopped letting agents say "done".
retained text
We run \~25 AI agent workspaces (PM → Dev → Tester → Critic → Release). Agents hand work off with markers in comments like PASS and RELEASE\_READY. We audited 8 weeks of it: \- Only \~9% of human interventions were real decisions. The rest was verifying, closing and re-routing work. \- 59% of tasks were closed by a human by hand, even though an agent had already said they were done. \- 8–11 per 100 tasks were false "done". For example, "30 files pushed" when nothing was on the remote, or a PASS below the critic's own threshold. \- The board-sweeping agent was 76% of all tasks. 92–95% of its runs ended with "no action". What we're trying: keep every agent as-is, but add a small service they talk to over MCP. \- Work contract per task type, listing what must be true: links resolve, UTM intact, the release package is on the remote branch, and so on. \- Checks run by the service on the pushed branch. Agents can't see or edit them, and some are hidden. \- Three states: PASS / FAIL / UNVERIFIED. Anything we can't prove is never faked as a pass. \- "Done" can only be set by the service. A marker from the wrong agent is ignored. We're starting in shadow mode, where it only watches and logs. Questions: 1. Has anyone measured how often their verifiers miss, e.g. by injecting known faults? 2. How do you handle subjective checks without an LLM judge becoming the authority again? 3. Is anyone doing continuous re-verification after "done"?
document fetched 2026-09-28 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wrqpyk/audited_1228_human_interventions_in_our_ai_agent