Context Windows

In one sentence

A context window is the maximum amount of tokenized content a model can consider in one request (and, in agent products, often across a managed session budget).

Why it matters

When context fills up, tools truncate, summarize, or drop history. Cost rises before quality collapses—then quality collapses.

How it works

• Treat the window as a hard capacity constraint set by the model/product.
• Everything in the window competes: system rules, files, tool output, history.
• Larger windows enable bigger tasks but raise the cost of careless inclusion.
• Monitor utilization; do not assume “max context” is free headroom.

Example

If the product allows 200k tokens and your session already holds 180k of files and logs, the next tool result may force eviction of earlier decisions the agent still “thinks” it remembers.