Agent kill switch

ProductionReliabilityPublished By Simon Budziak

An agent kill switch is an independent control that stops an AI agent from starting or continuing actions when risk becomes unacceptable. It must work outside the model's own reasoning, revoke active execution or credentials, preserve evidence, and leave systems in a state operators can inspect and recover.

NIST AI Agent Standards Initiative provides the primary reference used for this definition and its production boundaries.

How does agent kill switch work in production?

The control blocks dispatch, cancels workers where possible, and revokes credentials through the agent policy engine. It records state for agent recovery. A prompt telling the agent to stop is not a kill switch.

When does agent kill switch matter?

Trigger it during security events, uncontrolled cost, repeated policy breaches, or uncertain external effects. Agent incident response defines who may activate it, while an agent circuit breaker can stop known failure patterns automatically. Test the switch before an incident.

Frequently asked questions

What is agent kill switch used for?

Trigger it during security events, uncontrolled cost, repeated policy breaches, or uncertain external effects. Agent incident response defines who may activate it, while an agent circuit breaker can stop known failure patterns automatically.

Does a kill switch undo completed agent actions?

No. It stops further activity; rollback or compensating actions handle effects that already occurred.

Summarize this page with

Train your team to build this