2 Comments
User's avatar
Giving Lab's avatar

Really strong reframing from token cost to cost-per-completed-task.

One thing that helped our team: for each agent run, we log just 5 fields (goal, model route, tool calls, failure/recovery, next test). That makes throughput claims auditable and quickly shows where “cheap” setups silently leak time.

If useful, we share practical OpenClaw run breakdowns like this at https://substack.com/@givinglab.

Ari's avatar

Very cool! Thanks for sharing