> LLMs likely have a much better understanding of what our users need and want.

They don't.

Basically this sounds like Agentic Fuzz Testing. Could it be useful? Sure. Does it have anything to do with what real users need or want? Nope.