What if the agent fucks up the better approach but does a good job of the worst approach?

Then hopefully the reviewers will notice that the first prototype's flaws are correctable. Sometimes they won't, and they'll end up making a bad decision, just as they sometimes make bad decisions today with no prototypes to look at. But having prototypes allows for a lot of debates that are today vague and meandering to be reduced to "which of these assertions at the end of this integration test do you think is the correct behavior?".