Trust before the write, not after

Audit logs tell us what an agent did. What tells us it was right before it commits? For those piloting agentic workflows: where have you set the human-in-the-loop threshold, and on which transaction classes would you never remove it?