The context window is the total number of tokens the model can attend to in one call, spanning the system prompt, tool definitions, the accumulated transcript and the response. Everything the agent knows in a given step is inside it, and everything else has to be fetched.