Cost per agent run
Also known as Cost per agent run metric, Cost per agent run analysis
Definition
Cost per agent run is model expense incurred by one execution of an agent task. It gives teams a way to name, measure, or reason about an economic property of an AI workload without treating raw usage as proof of value.
+## Why Cost per agent run matters
Cost per agent run gives an AI program a more precise vocabulary than a single monthly invoice. Teams should define the measured unit, time window, owner, and decision the number is meant to inform. This entry focuses on model expense incurred by one execution of an agent task. Two workloads can have similar token counts while serving different users, routes, quality thresholds, or business purposes.
A concrete example
Consider this example: one run includes tool selection, intermediate calls, and a final answer. An analyst would record the relevant requests, separate direct model charges from adjacent operating costs, and state whether the figure is an average, total, or rate. The report should preserve dimensions needed to reproduce the calculation, including model, provider, environment, workflow, and outcome status where available. A useful review asks what changed, who owns the change, and whether the result remained acceptable for users.
How to use the measure
Start with a baseline, then compare periods with the same scope. Segment unusual results before changing a prompt, budget, or route. Pair the economic measure with quality, latency, reliability, and adoption signals so an apparent saving does not hide lower completion or more manual work. Keep pricing assumptions dated, document exclusions, and distinguish observed facts from forecasts.
Limitations
runs have different step counts, so averages need distributions. Cost per agent run should therefore be used as evidence in a decision rather than as a universal score. Recheck the definition when the workflow, model mix, billing rules, or success criteria change.
How this relates to Weave
For teams using Weave, Cost per agent run is most useful when connected to request-level token usage, model and route context, latency, quality signals, and workflow outcomes. That context helps teams investigate Cost per agent run by product or team and compare economic movement with engineering results. Weave data should support the measurement conversation, while the team remains responsible for defining the unit, validating assumptions, and deciding what action is appropriate.
Explore Token intelligence