The context window is everything the model can "see" in one request, counted in tokens: your instructions, any retrieved documents, the conversation so far, and the response it is generating. Run past the limit and something has to be dropped or truncated.
Why it matters in production
The window is a hard budget you spend on every call. Retrieved passages, tool results, and chat history all compete for the same space, and when they collide, the model silently loses whatever fell off the edge. Larger windows have made this less of a wall — but they turned it into a subtler trap rather than removing it.
The common mistake
Assuming a bigger window means you can stop being careful about what you put in it. Models exhibit a "lost in the middle" tendency — information buried in the center of a long context gets less attention than the same information at the start or end. Stuffing the window full of marginally-relevant text does not help the model; it dilutes the signal and raises cost and latency for a worse answer. More context is not more understanding.
How we actually use it
We treat the window as a scarce resource and spend it deliberately — this is the core of context engineering. That means reranking retrieved results so only the strongest passages make the cut, compressing history instead of appending it forever, and putting the most important instructions where the model actually attends to them. A big window is a bigger budget, not an excuse to stop budgeting.