"Undo" would work on the database. It doesn't work on your customer.

I always thought that the common answer to agent permissions is to let the agent act on anything you can undo and request only for crucial part. Sending a customer a messages usually would count as the thing you can't take back I guess.

I'm not sure that line holds for product decisions. In a past project, an agent just changed something that was easy to revert. But if we hadn't noticed, it could have set off a chain of follow-up actions that caused us big trouble under their own understanding. Another instance, when the AI agent marks a customer request as "planned." instead of Notice. It would be great if we found the issue early on, but if not, we might have missed this request

Meanwhile asking for approval with every changes would make things more secure, but when there are too many requests, I notice that people start approving without reading.

For people using agents around product decisions:

• Do you give actions like these their own permission level? Or do Yall just let it handle all task themselve.

• What would help your team review approvals more carefully? Fewer requests, more context, batching action, or something else?

• Has an agent made a call that sound right and turned out wrong? What did you change afterward?