AI Agent Kill Switch
An AI agent kill switch is a tested control that stops an agent from taking further actions — including work already in flight — while preserving a complete record of what it did. It works independently of the agent's own code, and restart requires a different person's recorded approval.
What an AI agent kill switch means
Most teams have a "disable" button; few have timed what happens after someone presses it. A real kill switch passes a five-step drill: stop new actions (revoke credentials or block at the gateway, never trusting the agent to stop itself), stop in-flight work (drain queues, cancel retries — zero actions after the stop decision), list everything the agent wrote (complete, with before/after values, in minutes), decide what to undo (every write has a rollback path or a named human owner), and restart under approval (blocked until a named approver — different from the person who stopped it — signs off). A deployment rollback is not a kill switch: it changes code but does not stop a running agent or account for its actions. For regulated deployments, the drill is the evidence supervisors and auditors ask for.
Frequently asked questions
How is a kill switch different from disabling an agent?
A disable button is untested and usually incomplete — it often stops new runs but not queued jobs, retries or in-flight work. A kill switch is proven by drill: zero actions after the stop, a complete write list in minutes, rollback paths for every write, and approval-gated restart.
Who should be able to trigger a kill switch?
A role on call — not only the agent's builder. Restart authority belongs to a named approver who is different from the person who stopped the agent. The separation is standard separation-of-duties, applied to agents.