Guide
Why AI agents use so many tokens
Agents resend context because each model call needs enough state to choose the next action. Tool output, conversation history, and retrieved documents can compound across a loop.
Tool calls become context
A terminal command or API response is useful once, but its full output may be carried through several later calls.
History is not free
Long conversations resend old decisions and intermediate material unless an explicit retention or compression policy exists.
Account by source
Break spend down by model, tool, MCP server, and cache status before choosing a reduction policy.