Agent step count measures the total reasoning or action steps an agent takes from task start to completion. It exposes accumulated execution complexity that single-call latency can hide. section 4.2.3
A rising count can indicate a more complex task, unexpected looping, repeated retries, or inefficient tool use. Interpret it with model calls, tool calls, loop rate, dependency failures, and total task completion time. section 4.2.3 section 4.6
Use the metric to investigate where an agent task accumulates work:
- Compare step count with model-call and tool-call counts to locate reasoning or external-system overhead. section 4.2.3
- Check loop and retry rates when the count grows unexpectedly; repeated steps can signal brittle integrations or unclear instructions. section 4.2.3
- Pair the metric with total task completion time because cumulative work, not model speed alone, shapes perceived reliability and latency. section 4.2.3 section 4.6
Cumulative task work
An agent task requiring five model calls, three retrieval calls, and two external tool calls may feel slow even when each individual call is fast. Step count helps reveal that accumulated execution burden. section 4.2.3
