← All articles
AI Agents3 min read

Your AI agent needs a kill switch

Every production agent eventually does something nobody expected. A kill switch is the difference between an incident and a disaster — here is what a real one looks like, and why you should test it like a fire drill.

Your AI agent needs a kill switch

Every production AI agent eventually does something nobody expected. A prompt change makes it confidently wrong at scale. A tool call starts looping. An integration hiccup turns a helpful assistant into a machine that refunds or messages customers on repeat. The question is never whether it will happen — it is how long it takes you to stop it. Most teams discover they have no way to stop it at all — except a fix deployed while the damage compounds.

A kill switch is not a feature. It is the difference between an incident and a disaster — and it is not the same thing as an alert. An alert tells you something is wrong; a kill switch stops the bleeding while you figure out what went wrong.

What a real kill switch looks like

First, it halts new agent runs immediately — not after the next deploy, but the moment the person on call flips it. If stopping your agent requires a code change, you do not have a kill switch. You have a deployment process, and during an incident that is the same as having nothing.

Second, it handles runs already in flight. An agent mid-conversation, or halfway through a multi-step workflow, should finish cleanly or cancel gracefully — never freeze mid-refund or leave a customer talking to a dead session. Decide in advance: complete the current step and stop, or cancel and hand over to a human. Either is fine. Discovering you never decided is not.

Third, it falls back to something safe without a deploy. When the agent stops, the work routes somewhere sensible: a human queue, a static response, or the old non-AI flow. A kill switch that leaves customers staring at an error page is barely better than no switch at all.

Fourth — the part teams skip — the switch must be operable by whoever is on call, not only by the engineer who built the agent. Incidents do not wait for office hours. If the only person who can stop the agent is asleep, the switch is decorative.

Test it like a fire drill

A kill switch you have never flipped is a theory. Run it in staging at least once a quarter: trigger it mid-conversation, time how long from "something is wrong" to "the agent is stopped", and check the fallback works and a human picks up the stranded work. Log every activation, including the drills, so the next incident starts with a record instead of a guess.

And keep the switch boring: one control, in one place, with one obvious effect. The moment it needs a runbook, a dashboard, and a Slack thread to operate, it will not get used when it matters.

We build production AI systems designed to fail safe — kill switch, evals, and fallback in place before launch, not after the first incident — because an agent you cannot stop is not a product. It is a liability with a chat interface.

Field note: A client's support agent started re-sending the same tracking update to a customer every few minutes after a Friday-evening prompt change. The only way to stop it was a deploy, and that needed the engineer who had written the change — who was offline. By the time it was stopped, the customer had received the same message fourteen times. The next version shipped with one switch the on-call lead could flip from their phone, routing everything to the human queue in under a minute. They have used it twice since. Both times the incident lasted minutes, not a weekend.

Related service Workflow automation

Building something worth shipping?

We take on a small number of AI product engagements. Tell us what you are building — we reply quickly.