Kill switch

A kill switch is a control that lets a person pause, cancel, redirect, or take over an agent mid-task, freezing or reversing its work the moment something looks wrong instead of waiting for the run to finish. The point is to stay reachable while the agent is still running, so a problem you spot at step three does not keep compounding through step thirty. Picture an agent told to clean up an email inbox by archiving old threads. Halfway through you notice it is archiving messages you still need. A kill switch is the button that stops it on the spot and, if the design allows, undoes the archives it already made, rather than leaving you to watch the rest of the run finish first.

Builder example

An agent with real account access can do damage faster than you can read its output, so the ability to interrupt it matters as much as the ability to start it. Imagine a scheduling agent that begins moving every meeting on your calendar after misreading one instruction. Without a way to halt it, you watch the mistakes pile up; with a kill switch wired into the loop, you stop the run at the first wrong move and review what it already changed. Tell the agent to check a stop flag between steps so an interrupt takes effect quickly.

Common confusion: A kill switch is not the same as a guardrail. A guardrail blocks a forbidden action before it ever runs, while a kill switch is the manual control you reach for when an allowed action is going wrong in the moment and you need to stop the run yourself.