THE ECONOMICS OF AGENTS

CHECK YOUR UNDERSTANDING • Why do input and output tokens get priced separately? ◦ Answer: Generating output is more computing work than reading input, so providers usually charge more for output tokens. • A database lookup takes 20 milliseconds and a model call takes 4 seconds. Where should you look first for savings? ◦ Answer: At the model calls, because inference is usually the dominant cost and delay. • Name three costs of running an agent that are not on the token bill. ◦ Answer: Any three of sandbox compute, observability storage, databases, evaluation runs, engineer time, and the cost of failures. • What is the purpose of a monthly FinOps review? ◦ Answer: To break spending down by team, agent, and task type, and to catch unexplained spikes before they become budget surprises. • Why should you measure total operating cost and not only the token bill? ◦ Answer: Because sandboxes, storage, databases, evaluation runs, and engineer time also cost money, and ignoring them hides the true price of running the agent.