Kill Switch

What is a kill switch?

Picture a self-driving car glitching on the highway. The last thing you want is a smarter steering wheel — you want a brake you can stomp. A kill switch is that brake for an agent: when it starts misbehaving, one press (or an automatic rule) makes it stop right now.

Why does an agent need one?

Agents overstep
An agent running tasks on autopilot might delete files, place duplicate orders, or blast a group chat. By the time it finishes, the damage is done.
Errors compound
Agent actions chain together, and one small mistake can snowball. The sooner you stop it, the smaller the loss.
Humans need a backstop
Even the best monitoring misses things. A manual "stop now" button is the safest last line of defense.

How to design one

Separate the permissions
Keep "stop" and "run" apart so a runaway agent can't switch off its own kill switch.
Multiple triggers
Support both a human hitting stop and rules that fire automatically — over budget, too many failures.
Make it recoverable
After stopping, you need to roll back safely and figure out what happened, not inherit a mess.

How it relates to a circuit breaker

A kill switch is more of a human backstop; a circuit breaker is more automatic protection. They usually work together — the breaker auto-stops first, and the kill switch is there when you need to shut the whole thing down.

Bottom line: a kill switch is the emergency brake you slam when an agent goes off the rails.