What is a kill switch?
The way an AI agent is stopped from outside itself: halt the run, disable the trigger, revoke the access.
Definition
kill switch (n.) – the mechanism that stops an AI agent's next action from outside the agent, without depending on the agent's cooperation, by halting the run, disabling what starts it, and revoking the access it acts through.
A kill switch for an AI agent is what ends a run that has gone wrong, from a layer the agent does not control. It is not one button. In this book's frame it has three parts: a halt that ends the run in progress, a disable that keeps the schedule or trigger from starting it again, and a revoke that removes the access the agent acts through. A switch that only halts leaves a scheduled job that starts again at the top of the next hour. A switch that halts and disables leaves a credential that still works for anything else holding it.
The term is used loosely. Vendors use it for a stop instruction in the prompt, a button in a dashboard, or revoking the agent's identity. Aembit defines an AI kill switch as "any mechanism that can stop, pause, or contain an AI system or agent" and says it "can be an instruction, a revoked credential, or a terminated process." The definition here is narrower: only a mechanism that works while the agent is wrong counts, which rules out the instruction. An agent that misreads its task can misread the order to stop.
A kill switch is part of the procedure, not an emergency accessory. Each procedure an agent runs states, next to its rung on the mandate ladder, how it is stopped: which process to end, which schedule entry to disable, which access record to revoke, and who may pull it. Under the runtime law, every automated job already runs by a named role, through a recorded access grant, on a defined schedule. The kill switch is those same records read in reverse. A job that cannot be stopped by disabling its schedule entry and revoking its grant was never registered properly.
The switch is tested before it is needed. A switch that has never been pulled is a belief about the system, not a fact about it. Pulling it once on a quiet day shows whether the run actually stops, whether anything restarts it, and how long the stop takes. That time is the gap between noticing a problem and ending it, and for anything that repeats on a schedule it multiplies the blast radius.
The limit is the one every honest definition states. A kill switch stops the next action and does not undo the last one. Aembit puts it plainly: "It can't reverse a payment that was already sent or a record that was already deleted." Repairing what already happened belongs to the recover step in the playbook. That is also why a kill switch does not replace a judgment gate: the gate holds an irreversible action before it happens, the switch ends a run after it has started going wrong.
Kill switch meaning, with an example
Asked plainly: if this agent started doing the wrong thing at 3 a.m., what ends it, and does that work without the agent's help? A kill switch is that answer written down and wired up. The bar is the same as for any document an agent runs: a person who has never seen the procedure can pull the switch from the procedure's own files, which is what cold-start-safe means applied to stopping.
A hypothetical: say an agent updates product prices from a supplier feed every hour, on apply within a cap. One night the feed arrives with every price in the wrong currency. The agent's verify step checks that each update was written, and each one was. The switch is three moves: end the running job, disable the hourly schedule entry so it does not start again, and revoke the pricing credential so no other script reaches the catalog through it. If only the first move is made, the job is back within the hour with the same feed. The prices already changed are then restored by the recover step, because the switch did not and could not touch them.
How kill switches fail
The switch lives inside the agent. A stop instruction in the prompt, or a stop tool the agent is asked to call, depends on the agent reading the situation correctly, which is the thing that has already failed. The switch belongs in a layer the agent does not control: the scheduler, the process, the credential, the provider account.
The switch closes one path and leaves another. Aembit notes that "a static API key stored in an environment variable, or a credential the agent already copied will keep working." Access that was never recorded cannot be revoked from the record, which is the practical case for org as data: revoking a role's access is deleting a line only if every grant is a line.
The switch is pulled and then quietly released. A procedure that was stopped returns to the propose rung until the cause is found and the playbook amended, not to the rung it held before. Restarting it unchanged repeats the run that needed stopping.