The point is the agent writes the specs and the human reviews and revises or asks for another rewrite that takes 90 seconds or less. So specs are both cheaper and better than anything I've seen before. They still get a lot wrong and it is hard to review them very carefully, but its easier than reviewing code to suss out design intent.
The point is that sometimes you don't know the spec is wrong until you've built the software and it's being used.