Context Window
The scarcest resource in agentic operations.
In one sentence
The text limit the model can see at once. It is the scarcest resource in agentic operations: everything that goes in (instructions, files, history) competes for that space.
Before
Everything that could go in
What it does
The window: what fits at once
After
What the agent effectively knows
Example
Ana’s field report is 2 pages long. The history from the previous three years is 40. The organization’s policies, 15. All of that competes for the same window, and whatever does not fit, the agent simply does not know. That is why deciding what to cut, which seems technical, is a business decision.
In practice
- Measure actual consumption per execution before optimizing anything.
- Separate what is fixed (instruction, policy) from what is variable (the case). The fixed portion becomes a candidate for cache.
- Define and write down what gets cut first when the limit is exceeded, so you are not relying on improvisation.
How to make it tangible
Measure: how many tokens your execution consumes, and how much of that is instruction, data, and history. A screenshot is enough.
Connects with
Translated from Portuguese with AI assistance.