Measure workflow usefulness, not agent count
AI workflow metrics should show whether the process improves evidence, cycle time, review quality, or recovery. Number of agents and generated text are activity measures, not outcomes.
Collect only signals that support a decision and state what the metric cannot prove.
| Signal | Possible question |
|---|---|
| Lead time | Did the bounded slice reach review sooner? |
| Rework | How often was the output returned? |
| Evidence quality | Can a fresh reviewer reproduce the claim? |
| False positives | How much review noise was created? |
| Recovery | Can an interrupted task resume from an artifact? |
Metric boundaries
- Define the workflow and comparison baseline.
- Choose a small set of observable signals.
- Record collection method and missing data.
- Review qualitative feedback with the numbers.
- Change the workflow only when the signal supports it.
- Avoid claiming causation from one task or one anecdote.
Measurement record
# Workflow measurement
Workflow: <name>
Baseline: <comparison>
Signals: <list>
Observed data: <facts>
Unknowns: <limits>
Decision: keep / adjust / stop
Status: documented example; outcome: Unknown.Canonical lesson source · JavaRush
Open JavaRush material