@agenticamit: Why AI agents burn through tokens An AI agent sends its model the entire conversation again on every single step, because the model keeps no memory between calls. My run sent 343,731 tokens that way to build one small web page, and 86% of them came from the cache, which on Claude Opus 5.5 costs 5% of the normal price. In Anthropic's own data, an agent typically uses about four times the tokens of a chat.