Most agent cost problems are not model-pricing problems. They are untracked-run problems: nobody can say which run spent what, for which task, and whether the output was used.
Here is the minimal tracking schema I use for every autonomous agent run. You can start with just a CSV.
| column | why it matters |
|---|---|
run_id |
lets you join logs, costs and outcomes |
agent_role |
which agent spent the money |
tokens_in / tokens_out |
the raw driver of cost |
cost_usd |
computed at write time, not at month end |
outcome |
shipped ,no_action ,failed ,retried |
cost_usd divided by the count of outcome = shipped. If it is rising, your agents are getting less efficient. outcome = no_action. This is pure waste, and it is often 20-40% of spend in loops that re-read state without acting.
import csv, time, pathlib
PATH = pathlib.Path('agent_runs.csv')
def log_run(run_id, role, tin, tout, cost, outcome):
new = not PATH.exists()
with PATH.open('a', newline='') as f:
w = csv.writer(f)
if new:
w.writerow(['ts','run_id','role','tokens_in','tokens_out','cost_usd','outcome'])
w.writerow([time.strftime('%Y-%m-%dT%H:%M:%S'), run_id, role, tin, tout, round(cost, 4), outcome])
Call it in a finally block so failed runs are recorded too. Failed runs are where the money hides.
I packaged this schema as a Notion database template plus a CSV starter and the three views above, so you do not have to wire it up yourself: Agent Ops Tracker (Notion + CSV) - https://renevibe76.gumroad.com/l/ptfvli (pay what you want, minimum $1).
The CSV approach above is fully enough to get started, though. The habit matters more than the tool.
What do you track per agent run? I am curious whether anyone measures cost per useful outcome rather than cost per call.