Cloud Assumes You Know What a Request Will Cost

Cloud infrastructure was built around a request whose cost you can estimate before it runs. You size timeouts to it, forecast capacity from its average, rate-limit against it, and bill by it. Agent workloads break that assumption at the source: an agent decides what work to do while it executes, so the count of tool calls, model invocations, and retries is only settled once the run is finished.

This session traces what fails when unpredictable work lands on infrastructure designed for predictable requests. Timeouts fire in the middle of legitimate long runs and surface as silent failures. Capacity plans built on averages miss a heavy tail that a few runs can dominate. Per-request billing and rate limits drift away from the resources actually being consumed.

The larger argument is that the request is the wrong thing to price and bound. The talk proposes a different unit of accounting, closer to the run and its budget, and covers what teams can do about timeouts, forecasting, and cost controls today while the platforms catch up.