Systems••6 min read
Measure Agent Work Beyond Speed
The signals that tell you whether an agent workflow is becoming more dependable, not just faster.
Track useful completion
Measure work that reaches a verified, usable state rather than counting generations or tool calls.
Watch rework
A fast first pass is less valuable when it creates repeated corrections. Track where work loops and which handoffs cause the most churn.
Keep humans in the metric
A good system improves judgment. Look at review time, confidence, and how quickly a person can understand and redirect the work.