Shut down an AI agent when nobody acts on its output. Not when it makes mistakes, and not when it is slow to find its feet, but when its work changes nothing: no draft gets sent, no analysis moves a decision, no task gets picked up by a colleague. MING Labs runs its production agent fleet on this criterion and used it to shut down its first agent after thirteen days.[S1]
01The signal: the Acted-On Rate
The Acted-On Rate is the share of an agent’s output that a named human actually uses or acts on. A sent draft counts. A forwarded analysis counts. A report that shapes a decision counts. Output that is technically correct and practically ignored does not count, no matter how much of it there is.[S2]
Activity metrics flatter agents. Message volume, tasks completed, and automation rate all measure motion, and an agent can score high on every one of them while producing nothing anyone touches. The Acted-On Rate is the one number in MING’s fleet reviews that separates a slow-starting colleague from noise.
The takeaway: judge the trend of use, not the volume of output. The two failure profiles look identical on activity metrics and opposite on the Acted-On Rate.
| Signal | Slow-starting colleague | Noise: shut it down |
|---|---|---|
| Output volume | Low to normal, uneven | High and steady |
| Acted-On Rate | Above zero by week two, climbing | Flat at zero |
| Human response | People correct and resend its drafts | People ignore its messages |
| Role design | Owned domain, named human owner | Capability without a domain |
| What to do | Keep it, tighten the role | Shut it down, redesign the role |
02What Major Tom taught us
MING Labs gave its first agent, Major Tom, a capability: coordination across the founder team. It never gave him a domain. Over thirteen days he sent 764 messages, and the team acted on none of them. Every message was technically correct and practically useless. We shut him down on March 30 and published the audit.[S1]
The lesson was not that the model was weak. The lesson was that the role was wrong: capability without accountability is noise. The agents that replaced him got job descriptions, owned domains, and a named human owner, and their output started being used within days. Shutting an agent down is not the opposite of adopting agents. It is how a fleet stays honest.
Nor was he unique. Across MING’s fleet, 15 to 20 percent of agents are decommissioned within their first quarter.[S2] That rate, why it is a sign of health rather than failure, and the retirements behind it have their own page: how often AI agents get shut down .
03Why week one tells you almost nothing
The obvious objection is that any new colleague needs time, and killing an agent in week one punishes slow starters. That is true for quality and irrelevant for use. Early output is allowed to be mediocre; a human editing a mediocre draft is still acting on it, and that counts. What week one cannot excuse is a rate that never leaves zero while volume climbs, because that pattern does not improve with polish. It improves with a different role.
The measure also has to sit outside the agent. An agent asked to grade its own usefulness will find itself useful, which is why verification is a job, not a setting . In MING’s fleet the Acted-On Rate is read from what humans did, never from what the agent reports about itself.[S2]
04The measurement, scoped
The founding data point: 13 days, 764 messages, 0 acted-on outputs.
Measurement scope
- Period: 17 to 30 March 2026, ended by shutdown.
- Sample: one internal agent (Major Tom, fleet coordination role) in MING Labs’ own production fleet; message audit n=764.
- Measured: whether any output was used or acted on by a named human.
- Success: output used at least once, rate climbing. Shutdown trigger: Acted-On Rate flat at zero across the window.
- Disclosure: scope only. Internal prompts, protocols, and per-agent tooling are not published.
Since then the same measure has governed promotions, not only shutdowns. Agents in the fleet earn autonomy levels as their Acted-On Rate climbs, each under a named human owner who must pass the Comprehension Obligation to keep the autonomy level up.[S2] Which work an agent may own in the first place is a separate question, and the ABC Framework answers it: judgment stays human, structured expert work is shared, routine gets owned.
