A kill switch and a budget do different jobs
A budget stops an agent automatically at a number you chose in advance. A kill switch is for what you didn't plan: you've seen something wrong and you want everything stopped now, whatever the budgets say. You need both, because a budget can be set too high, and a loop can do harm that isn't about cost.
What it has to do
- Sit outside the agent. A flag the agent is meant to check relies on the agent working properly, and a misbehaving agent is the reason you're reaching for the switch. The stop belongs in the path of the call, where the agent can't skip it.
- Take effect on the next call. Waiting for a deploy, or for a cached setting to expire, means more spending while you wait.
- Need no engineer. Whoever spots the problem should be able to stop it from a phone, without a terminal.
- Fail closed. If the switch can't be read, calls shouldn't go through.
- Say why. A paused agent should get a clear refusal, so it stops cleanly and doesn't retry.
- Be slower to undo than to use. Stopping should take one click. Resuming should ask for a second look, because it lets spending start again.
Calls already running
A stop at the check prevents new calls. It won't recall a request that has already left, and that request will finish and be billed. So expect a little spending after you press the switch, and don't read it as a failure. If your agents make long calls, cancelling in-flight work is a separate job for the agent's own runtime.
Decide the scope in advance
| Scope | Use it when |
|---|---|
| One agent | You know which one is at fault. |
| One key | One app or client is the problem, or a key has leaked. |
| One workspace or team | A group of agents is affected, and others must keep going. |
| Everything | You don't yet know what's wrong. |
When a key may have leaked, revoke it as well. Pausing stops spending, and revoking removes the access.
Test it before you need it
- Decide who may press it and who may undo it, and write that down.
- Press it during normal hours, with a test agent running.
- Confirm the next call is refused, and the agent ends its run without retrying.
- Confirm the people who should hear about it did.
- Resume, and confirm work starts again cleanly.
A switch that has never been pressed is an assumption. Repeat the test now and then, and whenever you change how agents reach the model.
Spend is in early access: you can set budgets and see what each call cost today, and budgets that stop calls are being switched on across the platform. The kill switch is one of the parts being switched on.
Questions
- What is an AI agent kill switch?
- A control that stops an agent's calls at once, without a code change. It sits in the path of each call, so it works even when the agent itself is misbehaving.
- Is a budget the same as a kill switch?
- No. A budget stops an agent automatically at a number set in advance. A kill switch stops it when a person decides to, whatever the budget says. Most teams want both.
- Does a kill switch stop calls already in progress?
- Usually not. It refuses new calls, and a request that has already been sent will generally finish and be billed. Cancelling in-flight work is a job for the agent's runtime.