Price the whole task, not one agent call.
An agent can make several decisions and tool calls before producing anything useful. Budget around a finished task so the invoice reflects the workflow you are actually buying.
Draw the practical limits
Set a ceiling on iterations, retries, wall time, and spend. Define what completion means and when the system should ask for help. A task that keeps running without progress should stop, not receive an unlimited token budget.
Include the supporting services
Add search, storage, sandbox execution, and external API charges. Track failures and manual review separately. Model usage is just one part of the cost when tools perform work on the agent's behalf.
Pilot before scaling
Test a narrow workflow with observable results and restricted permissions. Compare completion rate, cost, and human effort against an ordinary automated process. Credits can support a pilot, but a stable production budget needs predictable task boundaries.