A context window is the maximum amount of text a language model can take in and consider at one time, counted in tokens: pieces of words. Everything the model uses for a response, from its instructions to the documents and tool results, has to fit inside it. Anything outside the window, the model simply does not see.
Why does the context window matter once agents act?
An agent working on a long task keeps adding to its context: each tool call returns data, each step adds reasoning, the conversation grows. Eventually something has to go. If the system drops the oldest material first, the first thing lost may be the instructions or the policy that told the agent what it may not do.
The failure is quiet. The agent does not announce that it has forgotten a rule. It just stops applying it.
How does the window work in practice?
Every model has a fixed limit. The application assembles the context for each call: instructions, retrieved facts, history, tool results. If the total would exceed the limit, it must trim, summarise or leave things out. The choices it makes are part of context engineering.
Three practices help agents stay safe within the limit:
- Keep the agent's role and rules in a fixed part of every call, never in the part that gets trimmed.
- Fetch facts again when needed rather than carrying old copies forward, so the agent works from current records.
- Summarise earlier steps into short notes marked as the agent's own summary, not as source facts.
How agents carry what matters across longer gaps is covered in how AI agents remember context between tasks.
What does that look like in an example?
A commercial insurer's agent reviews a broker submission with a long schedule of properties. Partway through, the context fills. A naive setup drops the start of the conversation, which contained the underwriting guidelines. The agent goes on to recommend cover for a property type the guidelines exclude. A better setup pins the guidelines, processes the schedule in batches, and keeps a running list of exclusions found. The window was the same size in both cases; the design was different.
Is a bigger context window better?
Larger windows let an agent read longer documents in one pass, which helps. They do not make the agent pay equal attention to everything in them, and they do not stop it being misled by an instruction buried in page forty of an attachment. Bigger windows also cost more per call. Size is a capacity, not a control.
How is a context window different from a reversal window?
They share a word and nothing else. A context window is how much text a model can consider at once. A reversal window is the period after an agent acts during which the action can still be undone, such as the time before a payment batch is submitted. The first is about the model's input; the second is about the business consequences of the agent's output.
Questions people ask
Does a model remember anything outside its context window?
Not within a call. A model has no memory of earlier calls unless the application puts that information back into the context. Any longer-term memory is a store the application manages.
Can an agent's rules fall out of its context window?
They can if the system trims carelessly. Keeping rules in a fixed part of every call helps, and checking actions in a layer outside the model means a rule still applies even if the model lost sight of it. That is the role of the governed agentic operating system.