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