Measure whether the agent finished the job.
ATCR replaces pageviews and fragmented sessions with the production outcome that matters: a completed agent task.
Canonical definition
Agent Task Completion Rate
completed production agent tasks ÷ production agent tasks started
Classification
Start, complete, block, resolve.
Started
Evidence shows a concrete task entered an observable product path.
Completed
The defined production outcome was observed.
Blocked
The task reached a barrier and did not complete.
Unresolved
Evidence cannot support a reliable outcome classification.
Worked example
Keep unresolved evidence visible.
Illustrative reconstruction
Of 20 started production tasks, 12 completed, 5 were blocked, and 3 remained unresolved. Production ATCR is 12 ÷ 20 = 60%. The unresolved share remains visible; it is not silently converted into success or failure.
| Metric | Source | Answers | Limitation |
|---|---|---|---|
| Production ATCR | Production journeys | Did the real observed task complete? | Requires an observable start and outcome. |
| Eval pass rate | Controlled Eval runs | Did comparable test runs pass? | Does not substitute for production ATCR. |
| Evidence confidence | Reconstruction links | How well supported is this connection? | Does not measure completion. |
Measure the real outcome.
Start by reconstructing one task with a clear completion criterion.