Context window

A context window is the amount of text an LLM can hold in a single call — everything it can "see" at once, including instructions, conversation history, and any documents provided.

Every model has a hard ceiling on how much it can read in one go. Once a conversation or a document set grows past that ceiling, older content has to get summarized, dropped, or handled a different way — which is why AI tools sometimes "forget" something you told them earlier in a long session.

Subagents each run in their own isolated context window — a bounded task gets delegated out, does its work, and only its final result returns to the parent run, so the parent's own context never fills up with a subagent's intermediate reasoning.