Measure tokens, tool calls, and latency during agent evaluation

Goal

Create comparable baselines for tool calls, tokens, and runtime across agent iterations.

Steps

1. Choose a fixed evaluation set and run configuration; record model and SDK versions. [efficiency]
2. Read tool_call_count_v1, inference_call_count_v1, token_usage_v1, and invocation_duration_v1. [efficiency]
3. Compare quality and resource metrics before and after changes, noting sample size and measurement limits. [efficiency]
4. For clear regressions, optimize tool planning or context and rerun the same benchmark. [efficiency]

How to verify

The report shows quality, model calls, tool calls, tokens, and latency under matching settings.

Troubleshooting

If usage is not comparable, align cases, concurrency, and model version; efficiency metrics alone do not establish quality.