The formula is simple: completed tasks divided by eligible tasks. The difficult part is defining completion before the run. An agent that opens a support ticket, drafts a reply, and sends it has completed three different things depending on the actual goal.
What makes a completion metric credible?
Define the goal, evidence, exclusions, and treatment of partial success in advance. A metric with no clear success rule only counts the outcomes a team wanted to see. AI agent evals should use representative tasks, including cases where the correct result is a refusal or a human in the loop handoff.
Why should it not stand alone?
An agent can complete work through unnecessary calls, unsafe steps, or an expensive retry loop. Completion measures the destination, not whether the route deserved trust. A completion rate needs companion measures before it can justify autonomous action. Pair it with agent trajectory checks, trace-based evaluation, cost, latency, and business outcomes such as AI ROI measurement.