Design the Pause, Not Just the Permission Rule

Sally’s example shows how an accurate answer can still miss the detail that changes someone’s next decision. Zaidan asks where an AI agent’s approval boundary belongs, especially when small actions can add up to consequential ones.

My third perspective: a pause is an interface state, not merely a policy rule. When an agent stops, the person should be able to see why it stopped, what remains safe to do, who owns the decision, and what will let the work resume. Otherwise, “approval required” is a locked door with no sign explaining what is behind it.

Prototype three moments for one realistic task: the agent is preparing, approval is needed, and work can resume—or must stay on hold. Ask someone to use the flow without coaching. Where do they hesitate, or mistake a request for approval as permission to act?