As enterprises move from single LLM calls to autonomous agents, spend gets harder to see and control — a single agent can fan out into dozens of model calls and tool invocations, with no clear owner of the bill. I'm the Senior PM on Harness's agent-cost product: tracking token and inference spend by agent, model and provider, attributing every dollar down to the individual agent, session and trace, and letting teams set budgets and hard spend limits by user, agent, team or business unit.