Token budget
A token budget is a hard ceiling on how much context is allowed into a request, enforced before the request is sent rather than negotiated with the model.
The distinction that matters is between a ceiling and a preference. A preference is something an agent exceeds when it feels underinformed, which is always. A ceiling is a number that context is assembled to fit, with whatever did not fit reported rather than silently dropped.
Budgets work because nothing inside an agent loop ever concludes that it now knows enough. Given an unbounded window, an agent will keep reading, and every file it reads is paid for on every subsequent turn of that session.
The exact number matters less than having one. A budget that forces a choice produces a different and usually better selection than no budget at all.