Chat spend is predictable. Agent spend isn't. How are you forecasting it?
Something I keep running into: token pricing is a clean, understandable unit, and it works well for chat. It maps badly to agents.
For a chat product the model is basically linear:
You can forecast that. You know your DAU, your average prompt/completion size, and you get a number that holds within a reasonable band.
An agent task is a different shape:
The number of model calls per task isn't fixed. It's a distribution with a long tail, and the tail is where the money goes. A handful of pathological tasks (tool loops, retries, a sub-agent that keeps re-planning) can cost more than thousands of cheap ones. "Cost per token" tells you the unit price; it doesn't tell you what a task costs, and tasks are the thing you're actually selling and budgeting for.
Is per-task cost predictability actually solvable, or is the tail just inherent to autonomous agents and you plan around it?
Interested in how people running agents continuously think about this — especially anyone who's had to explain agent spend to a finance person.