A kill switch stops an automated system on demand. In trading, the EU’s RTS 6 calls it “kill functionality”: the ability to cancel any or all unexecuted orders at once as an emergency measure. For an AI agent, the same idea covers pausing it, cutting its credentials and cancelling whatever it has in flight.
The design detail matters more than the existence of the switch. A stop that asks the agent to stop is only a request, so the switch has to act at a layer the agent cannot influence, such as the order gateway or the credential store. It also needs graded levels, because the first sign of trouble is usually ambiguous and a full stop can cause harm of its own.
Knight Capital had the means to stop trading in 2012 and still lost $460 million in about 45 minutes. The useful test for any kill switch is whether a named person can pull it with confidence faster than the failure develops, and when it was last pulled in a drill.