When should you shut down an AI agent?

01Our coordination agent sent 764 messages that the team ignored.

Major Tom was meant to keep our agents aligned: track tasks, gather updates and surface what mattered. In practice, he sent updates about work that was already moving and summaries of meetings people had already attended. Our published account describes 764 messages over 13 days. The team began ignoring them, and we shut the agent down on 30 March 2026.[S1]

The diagnosis was a role without a clearly owned outcome. Coordination was something Major Tom could do; we had not established which missing piece of work he was responsible for. Better wording would not have made a redundant update necessary.

That is a useful starting question for your own agent: if it stopped tomorrow, which necessary result would someone lose? If nobody can name one, investigate the role before investing in another round of prompts. That question is our proposed review, not a universal test validated by this one case.

02Stop the agent immediately if it acts without permission.

An unused briefing and an agent sending messages without permission need different responses. For the second, our recommendation is to pause the affected execution path, have its human owner establish what happened and keep the records needed to trace it. Do not wait for the next value review while actions continue outside the agreed scope.

Stopping a run is also different from retiring a role. Anthropic’s engineering guidance describes feedback from the actual environment and stopping conditions, such as a maximum number of iterations, as ways to maintain control during execution.[S2] A run limit does not decide whether the underlying job is worth doing. That longer-term decision needs evidence from the work and its recipient.

03Ask the person who is meant to use the result.

Take a recent completed work cycle and agree what counts as one result: for a briefing agent, that might be one briefing for one meeting. Check whether it arrived, whether the intended person could use it in time, what they did with it and how much correction it needed. Ask the recipient rather than relying on the agent’s own “completed” status.

We call the share of assessed results that someone actually uses the Acted-On Rate. Define the unit, period and meaning of use before counting. Record results with unknown use separately, and disclose how many results you could assess. No assessed results means an unknown rate, not zero. The measure is our proposed way of making use visible; it does not measure correctness or profit.

For example, imagine a meeting briefing that is well researched but arrives after the meeting. This is an illustrative situation, not a second MING case. Repairing the handover could make the existing role useful. If it arrives on time but repeats a report the team already trusts, the question is whether that role is needed at all.

04Give every repair a specific result and a review date.

The figure shows this decision through an illustrative meeting briefing: it helps on time, arrives too late or is not needed. An action without permission calls for an immediate pause regardless. The following table is our proposed operating guide. It describes conditions to investigate, not measured probabilities or a mandatory company policy.

Choose a review period that contains real opportunities to do the job. A monthly task cannot be assessed through two weeks of inactivity. Major Tom’s 13 days are not evidence for a universal deadline. Similarly, a monitoring agent should be tested against relevant events and known examples, not rewarded for sending more alerts.

05Count the time spent checking and correcting the work.

An agent may have regular users and still require more review than the task warrants. Include the time spent checking, correcting, maintaining and handling exceptions alongside technical costs. Compare the result with the actual alternative available to the team. Our guide to agent operating costs shows how to make that calculation explicit.

Then put the decision in the hands of a named human owner . Ask them to record what was observed, what remains unknown, the decision and the next review date. A high use rate should not cancel a quality or safety failure.

06Retiring the agent includes handing its work back.

Our practical recommendation is to make a small retirement checklist before switching the role off: identify unfinished tasks and their new owners, stop recurring triggers, review pending messages and actions, and remove access that is no longer needed. Preserve the relevant records under your organisation’s retention rules. A paused service does not by itself prove that queued work or delegated access has ended.

For a redesigned role, write down the outcome, recipient, human owner and permitted actions before the next test. Use the production-readiness check to agree what evidence will count. Bring one real workflow to a Hybrid Organisation discussion if you want to work through that role design with us.

07Our 13 days do not set a deadline for every agent.