How do I implement an AI agent kill switch?

An AI agent kill switch is a deterministic mechanism that immediately halts all agent execution — not by asking the model to stop, but by revoking execution permissions at the governance layer so no further actions can commit.

Most "kill switch" implementations are broken because they rely on the model obeying a stop instruction. But the model is the problem — you can't ask the thing that's malfunctioning to stop malfunctioning.

A real kill switch operates at the execution boundary, not the inference layer:

• Namespace freeze: Revoke all execution tokens for a specific agent namespace — pending commits return 403
• Global halt: Switch the policy engine to DENY-ALL mode — every evaluation returns FORBIDDEN
• Selective freeze: Block specific action categories (db.write, api.call) while allowing reads
• Automatic triggers: Configure rate-based kill switches that activate when an agent exceeds N actions per minute

Exogram's kill switch operates at Gate 0 (namespace validation). When activated, the agent can still reason — but every proposed action is rejected before evaluation. The model continues to think. Nothing executes. Recovery is instant: re-enable the namespace and execution resumes from the last committed state.