Execution is a revocable permission, not a default.

Agents that cannot stop themselves

Every existing autonomous agent executes whenever it is capable of executing. If the model can generate an action, the action is taken. There is no internal signal that says "I can act, but I should not act right now." Confidence, when it exists, is a post-hoc score attached to outputs — not a pre-execution gate that determines whether the agent should act at all.

This produces systems that are maximally active and minimally cautious. An agent experiencing rapid environmental change, degraded input quality, or internal state inconsistency continues executing at full speed because nothing in its architecture tells it to slow down. The result is confident-looking behavior from an agent that has no structural basis for confidence.

Confidence governance computes execution readiness from multiple inputs: internal state consistency, environmental stability, affective state, integrity coherence, and capability assessment. When confidence drops below threshold, the agent enters a non-executing cognitive mode — it continues to reason, monitor, and plan, but it does not commit actions. Reauthorization requires sustained confidence recovery above a hysteretic threshold, preventing oscillation between executing and non-executing states.

Safe autonomy through structural caution

The ability to not act is the most important safety property an autonomous system can have. Every catastrophic autonomous system failure involves an agent that continued executing when it should have paused. Confidence governance provides the structural mechanism: a multi-input computation that gates execution, a non-executing mode that preserves cognitive function, and a hysteretic recovery that prevents premature re-engagement.

This is not a kill switch. The agent is not shut down. It continues to think, observe, and prepare. It simply does not act until it has structural justification for doing so. Rate of change detection provides early warning: if confidence is dropping rapidly, the agent can transition to non-executing mode before reaching crisis, rather than continuing at full speed until failure.