What Agentic AI Actually Costs to Run

Token spend is the number everyone quotes and the one that matters least. The real cost of an agent in production is the evaluation harness around it.

What does an agent cost once it is actually doing work?

Most cost estimates for agentic systems stop at the model bill. That is the easiest number to obtain and the least useful one to plan against.

The bill nobody forecasts

An agent that takes real actions against real systems needs an evaluation harness, and the harness is where the engineering time goes. Every tool the agent can call needs a test that proves it behaves under failure, not just under success.

Why the token bill misleads

Token spend scales with usage, which means it is visible, predictable and easy to cap. Evaluation cost scales with the number of things the agent is allowed to do, which is a design decision made long before anyone looks at a bill.

What to do instead

Budget for the harness first. Decide what the agent is permitted to do, write the failure tests for each of those actions, and only then estimate tokens. An agent whose blast radius is small is cheap to evaluate and cheap to run.