But if you never review it, how do you know it’s correct, secure, maintainable, performant, etc? You just trust it’s correct because you told the agent to make it correct?

I guess so, more AI is always a solution. I still like to review what the agents produce from time to time.