I use coding agents every day. Most of what I paid for wasn't code; it was files it re-read every few turns and conversation history it kept carrying long after anyone needed it.

The fixes do not come from clever prompting. Anthropic reported that putting tool definitions behind a search tool cut those tokens by about 85%. Moving MCP tool calls into code execution took one workflow from roughly 150,000 tokens down to 2,000. Clearing stale tool results cut token use by 84% in a 100-turn evaluation. And prompt caching bills the unchanged part of a prompt at a fraction of the normal input price.

We will take a journey through real code, real research. We'll look at where we can save tokens and what it costs you in accuracy or complexity. You'll leave with a deeper understanding of your agent's token usage and a clear first change worth making.