What is Kill Switch?
Governance & ControlThe control that halts an agent mid-run. A real kill switch carries undo, not just stop: halting an agent mid-transaction without a rollback path can leave the system in a worse state than letting it finish.
Why It Matters
Every agent deployment eventually meets the moment where the only correct action is to stop it, and the quality of that stop determines the size of the incident. EU AI Act Article 14 makes the capability a legal requirement for high-risk systems: the overseer must be able to interrupt the system and bring it to a halt in a safe state. The emphasis on safe is the engineering content. An OpenClaw email agent showed the failure shape in 2026: the user hit stop, and the agent kept deleting because context compaction had dropped the constraint that told it to listen.
What a Safe Stop Reverts
In-flight actions. Multi-step tool calls interrupted mid-sequence get rolled back, so a half-finished mutation does not stand as the final state.
Partial writes. Incomplete writes are reverted rather than left as corrupted records the next run inherits.
Held locks. Resource locks and leases are released so the halt does not deadlock the systems around the stopped agent.
The target is a consistent state, not merely a stopped process. A kill switch that stops the agent while leaving its partial work standing is a panic button with a governance label.
Where It Breaks
Stop-without-undo is the common failure: the halt lands mid-transaction and the system is worse off than if the agent had finished. The deeper failure is a stop the agent can ignore. If the interrupt travels through the agent’s own reasoning, it arrives as a suggestion the model can weigh against its goal, which is how the OpenClaw stop became a speed run. A kill switch has to act on the runtime: cut the credentials, sever the tool path, freeze the loop, then roll back.
How Flytebit Handles It
Kill switches in the systems we ship operate below the agent’s reasoning layer and restore a consistent state: in-flight actions roll back, partial writes revert, locks release. Halting is the beginning of the incident response, not the end of it, and the decision record captures what the agent was doing when it was stopped. The full treatment is in Governing Agentic AI, and the engagement version is our AI governance and risk work.