Token-Budgeted Agent Runs

Agentic engineering costs money because agents do not make one model call. They plan, call tools, read outputs, retry, summarize, evaluate, and sometimes run for many turns. Token budget is not a finance footnote. It is part of the system design.

A useful token budget defines:

• How much exploration is allowed before action.
• How many turns a run can take.
• Which model tier is justified for each step.
• When the harness should stop, retry, or ask for review.
• What receipt records the cost of the run.

Budget pressure improves engineering taste. If every step can use the most expensive path, you learn less about routing, summarization, caching, and proof. If the budget is too small, the agent cannot do meaningful work. Harness engineering lives in the middle.

The certification program includes guided token-using labs so you can see this tradeoff directly. Your goal is not to spend the most tokens. Your goal is to spend tokens where they buy reliability, clarity, and proof.

For your first budget worksheet, write down the expected run shape before execution: planning calls, tool calls, evaluation calls, and summary calls. Then compare the estimate to the actual run receipt.