Recognizing the Limits and Taking Back Control

You've been in the session for forty minutes. The first attempt didn't compile, the second compiled but broke two tests, the third fixed those two tests and broke a third, and the fourth just reintroduced the error from the first attempt. You rerun it anyway, because "it's almost there" and because forty minutes is too much to throw away. An hour later, the file looks like nothing you recognize, you no longer know which part of the code is original, and you end up with an angry git checkout that undoes everything, including the two changes that were good.

This scenario isn't an anecdote. It's the dominant failure mode of working with an agent, and it comes down to a well-documented bias: the sunk cost fallacy. Each iteration increases your emotional investment and your reluctance to abandon it, while the probability of success keeps dropping. An agent that has failed three times at the same task isn't learning: it's spinning in a context polluted by its own failed attempts, which steer it toward the same dead ends. The skill to acquire is therefore counterintuitive: you have to know how to decide to stop, early, based on objective signals rather than gut feeling.