Kill switch
- Industry term
A kill switch is a control, kept separate from the agent, that stops an agent or workflow from starting more work and cuts off its access to tools and accounts.
A “stop” sent to an agent reaches it as one more input to weigh, so a looping or confused agent can carry on past it. A kill switch acts one level down, in the accounts the agent uses and the service that starts each of its runs, so it works without the agent's cooperation.
Where you’ll meet it
You set up an agent in your team's chat app to post one reminder each morning. After a settings change, it posts the reminder every minute. You type “stop” to it, and the reminders keep coming.
What to do
Go around the agent. In the chat app's admin settings, find the list of installed apps or integrations and disable the agent, or ask your workspace admin to. Once things are quiet, write down where that setting lives, so the next stop takes seconds.
Easy to mix up
| Kill switch | Approval gate |
|---|---|
| Any time, to halt work already running | Before an action, which waits for your OK |
Check yourself
You flip the kill switch on an agent that was partway through emailing invoices to clients. Before you write to any client, how do you find out which invoices went out?
Go deeperHow fast a stop takes hold depends on what it cuts
A stop flag asks the workflow to cancel between steps, so a long step runs to its end first. Revoking the agent's credentials, its keys and logins, refuses it new access, though access tokens, the short-lived passes a service already handed it, may stay valid until they expire. Shutting down the program that runs the agent halts it outright.
Requests the agent sent to other services just before the switch may still complete afterward, so measure how long a full stop takes on your platform before you need one.
In AI safety research the term raises a harder question, whether a capable system would resist being switched off. In a 2015 paper titled “Corrigibility,” Nate Soares and colleagues described the wanted property, a system's willingness to accept correction and shutdown. Research since then asks how to build agents with no reason to block their own off switch.