THE ECONOMICS OF AGENTS

TOKEN COST VERSUS TOTAL OPERATING COST The token bill is only one line in the real cost of an agent. Total operating cost is everything it takes to keep the agent running in production. Consider what else is involved. Sandbox and compute cost: Many agents run code or tools inside a sandbox, which is a locked, isolated computer environment where mistakes cannot damage anything important. Sandboxes use computer time and memory, and those are billed by the minute or the hour. Observability cost: Observability means being able to see what your agent did, through logs, traces, and metrics. Recall from Chapter 28 that a trace records an agent run step by step. Storing and searching millions of these records on an observability platform costs money, and agent traces are large because they contain full prompts and responses. Storage and data cost: Memory stores, document indexes, and databases the agent uses all have monthly fees. Evaluation cost: Running test suites against the agent calls the model many times, so testing itself consumes tokens. Human cost: Engineers spend time maintaining the harness, reviewing flagged outputs, and responding to incidents. People are often the most expensive part of the system. Failure cost: When an agent makes an error, there is the cost of refunds, rework, or lost trust. Think of owning a car. The price of fuel is easy to see, but the real cost of ownership also includes insurance, repairs, parking, and the time you spend driving. If you budget only for fuel, you will be surprised. An agent's token bill is the fuel, and the rest is everything else about owning the car.