canonical-url: “https://hunterpaulson.dev/agent-harness/tokens/” description: “Hunter Paulson’s personal website for writing, art, projects, and experiments in computer science.” og-type: “website” site-name: “hunter paulson” social-description: “Hunter Paulson’s personal website for writing, art, projects, and experiments in computer science.” social-image-alt: “ASCII black hole animation from Hunter Paulson’s personal website.” social-image-height: 769 social-image-type: “image/gif” social-image-url: “https://hunterpaulson.dev/assets/social/home-blackhole.gif” social-image-width: 769 social-title: “hunter paulson” twitter-card: “summary_large_image” —

tokens

At the highest level tokens can come from two sources: the model and the environment.

note: i am not calling these input and output because input tokens from previous turns contain messages, and therefore tokens that were generated by the model. if you are smart you know about prefill where you fake assistant messages but this is not used very often.