Everything a model knows about your task at the moment of generating has to fit inside the window: the instruction, the reference material, the conversation so far, the tool results. When it does not fit, something gets dropped, and what gets dropped is rarely what you would have chosen.
In creative automation this is the constraint that shapes architecture. A workflow that pastes an entire brand guideline into every call is spending most of its window on material that is irrelevant to the current step, which is why retrieval exists.
Bigger windows have not removed the discipline. They have moved the failure from "it did not fit" to "it fit and the model attended to the wrong third of it".
Does a bigger context window mean I can stop being selective?
No. Relevance still beats volume: a model given eighty pages will attend to some of them, and you do not choose which. Curating what goes in remains the highest-leverage thing you control.
How does this affect a creative automation build?
It decides whether you can paste reference material into every call or need retrieval. Most real systems need retrieval sooner than they expect.