How do you control what an AI agent is allowed to do, on whose behalf, at the moment it acts?

When should a person decide, and how do you make that decision hold up later?

Can you prove why an AI outcome happened, and that it was the outcome you allowed?

How do you make agent work survive failures, waits and retries without doing anything twice?

What is an agent allowed to know when it decides, and is that information current?

What does the software underneath production AI agents need to look like?

Why do agents stall between pilot and production, and what should you build versus standardize?