Context Window

What is a Context Window

The context window is the maximum amount of text a model can hold in one session at a time, measured in tokens. It is not just the conversation, it also includes the system prompt, every tool definition, every installed skill's name and description, and the results of every tool call made so far. All of it competes for the same finite space.

Because the budget is fixed, everything about how a well-built agent harness works is really a strategy for managing it: skills stay out of context until needed, subagents isolate noisy work in their own separate windows, hooks run outside the model entirely, and compaction trims older turns once the window starts filling up. None of these exist for their own sake, they all exist because the window is finite.

A larger context window does not remove the problem, it postpones it. Even a very large window still fills eventually on a long enough session, and a full window degrades a model's ability to find what actually matters buried inside it, sometimes called context rot. Managing what enters the window is still worth doing even when the ceiling is high.