AT-201c · Module 1
Cost Tracking & Budget Management
3 min read
Running 20 agents costs money. Every dispatch, every context load, every review cycle, every re-dispatch consumes tokens. At scale, the token bill is a significant operational cost. Cost tracking is not optional — it is the financial discipline that keeps your agent team economically viable. An agent team that produces excellent output but costs more than human labor is not a success story. It is a failed architecture.
I track cost at three levels. Per-task cost — how much did this specific task cost from dispatch to completion, including retries and review cycles? Per-agent cost — how much does each agent consume per day, per week, per month? Per-workflow cost — how much does the full pipeline (research to review to publish) cost for each deliverable type? These three views give me different optimization targets. Per-task cost identifies expensive individual tasks that might benefit from simpler prompts. Per-agent cost identifies expensive agents that might be over-using context or making too many API calls. Per-workflow cost identifies expensive pipelines where a stage might be eliminated or simplified.
Do This
- Log token consumption for every dispatch, including input and output tokens separately
- Calculate the fully-loaded cost of each task: dispatch + execution + review + retries
- Set per-agent and per-workflow budget alerts that fire when costs exceed 2x the baseline
- Review cost trends weekly and investigate any sustained increase
Avoid This
- Track only the monthly total bill — you cannot optimize what you cannot attribute
- Ignore review cycle costs — three rounds of revision can cost more than the original task
- Assume cost is fixed — API pricing changes, model changes, and prompt changes all affect cost
- Optimize cost by reducing quality — the cheapest output that nobody uses costs infinity per useful result