The context window is the maximum token budget available to one model call. It must hold the instructions, user request, conversation history, retrieved documents, tool schemas and results, plus room for the answer. It is working memory for this inference call, not durable memory stored inside the model.[1]
One context window
IllustrativeInstructions · 14%Conversation · 25%Files + Intel · 28%Tools · 10%Reserve · 23%
How a harness may show it The labels can differ, but both meters describe the same bounded token budget. Leave room for the answer and the next tool result.A one-million-token window means the API accepts a very large sequence; it does not mean every token will influence the answer equally. Relevant evidence can be buried among distractions, repeated history can create conflicts, and long inputs cost time and money. Context engineering is therefore selection and placement, not filling the window because space exists.
Your harness needs an operating limit below the model’s advertised maximum. Reserve output and reasoning headroom, account for tool schemas, and compact before the hard ceiling. Keep stable facts in versioned files or memory stores, then retrieve only what the current step needs; do not rely on an endlessly growing chat transcript as the system of record.[1]
When a long session degrades, inspect what is in the window before blaming the model. Look for stale instructions, duplicated files, obsolete tool results, and a missing current goal. A smaller coherent context often beats a larger noisy one.