Accepted Agent Throughput Scorecard

Running workers are not output. This measures accepted results per worker-hour, cost per accepted result, and whether your instrument can see your workers at all.

← the article

Can a worker slip between two of your samples?

Do this first. If a sampler reads the worker count on a timer, a worker that starts and finishes between two samples is recorded as zero, and every per-worker-hour number downstream becomes a lower bound rather than a measurement. No polling interval is safe for arbitrarily short workers; only per-worker start and end events are.

How often the worker count is recorded.

Not your timeout. The shortest job whose time you need counted.

Window A

One measurement window: a run, a day, a sprint. Leave a field blank if you did not measure it. Blank stays unknown and never becomes zero.

Window B, for a before-and-after

Fill this in to compare two worker counts on the same task family. The verdict is on accepted output per worker-hour, not on activity.

Verdict

Fill in both windows to get a verdict.

Measured reference: Demonblade Labs Issue 001

One real production system, five days, 340 instrument samples, 89 reviewed tasks. Measured, not synthetic. Observational, not a controlled experiment: the governor chose the worker count, task difficulty varied, and 85% of accepted results fell in intervals the sampler recorded as zero workers. The All row (25.2) is therefore a lower bound, not the system cost, and the 166.6 figure applies only within observed intervals. The true system-wide figure is unknown.

Lane-minutes per accepted result by concurrent worker count. Window 2026-10-17 07:47 to 2026-10-22 04:48 (+10:00). Method, dataset and limits: the article.
Concurrent workersWall hoursWorker minutesAcceptedNot acceptedWorker min / accepted

Definitions, so nobody argues later

Attempted
tasks dispatched
Accepted
an independent reviewer said ACCEPTED
Worker hours
summed active worker time, not wall clock
Attributable accepted
accepted results inside intervals where worker time was observed
Attribution coverage
attributable / accepted. Below 1, your cost per accepted is a lower bound, and the gap is not a correction factor

Limitations