An LLM sees one stream of tokens. All the different components—system prompt, user messages, tool results, model responses—serve different purposes, have different trustworthiness, and carry different authority, yet they all appear in one token stream.