Why agent costs scale wrong
The overhead before the work.
The field note
A run we estimated at about $15 cost $248. Seventy-two skills loaded into the researcher’s context on every call. The scorer wasn’t using prompt caching, there was no hard provider spend cap, and the cost monitor wasn’t running. Several small decisions added up to one expensive run.
Measure the actual request
Tool definitions, conversation history, and retrieved material can consume context before the requested work begins. The amount depends on the integration and loading strategy. Trace what is actually sent, and attribute definition tokens to each call.
Load only what the call needs
Progressive disclosure makes tools discoverable without placing every definition in every request. The homepage example compares 72 loaded definitions with 3 needed definitions. The example request needs three skills; the rest stay available to discover.
Put a boundary around the bill
Run a small sample and measure the actual cost before scaling. Set a hard spend cap, check caching on repeated inputs, and deliberately trigger an alert to prove the monitor works. Keep straightforward batch work in a resumable script when it doesn’t need an agent making decisions.