> "someone needs to say if it's been done"

If that's the end user, they're not happy. End users generally don't like doing unpaid involuntary QA work.

Customers are notoriously bad representatives for 'what to do', per "my customers would have asked for a faster horse". And they also hate being used as QA; most quality issues (which can be reliability, UX, latency, etc) need to be addressed before the customer sees it and forms their durable impression of the software.

And then also it's as you say, customers do need help from domain experts (whether they admit it or not).

What if customers have conflicting requirements? Or they change their mind midway through implementation? IMO agents are bad at handling ambiguity, and often make assumptions and tradeoffs that are lost after the conversation. This is why there's a lot of compounding technical debt in projects where humans don't inspect the code.