Engineering culture has strong norms about compute: you profile before optimizing, you don’t ship an O(n²) loop on the hot path, you know what your queries cost. Context windows have none of that discipline yet. Prompts grow by accretion, nobody owns their size, and the bill arrives as inference cost and degraded attention.
The mental shift is small: a context window is a budget, and every span in it is spending. Tool definitions, system boilerplate, stale history — each line should justify itself the way a dependency or a database call does.
Budgets need meters. Our field study of production MCP traffic found most requests spend the majority of their tokens on content with no measurable effect on output — waste you can only see once you instrument attribution. Measurement turns prompt hygiene from an aesthetic argument into an engineering one.
Budgets also change design. When context is treated as scarce, you scope it — the way W0rkTree serves agents a view of just the files a task touches, or the way structured protocols replace re-sent boilerplate with a schema the model already knows. Scarcity produces architecture; abundance produces sludge.
The models will keep getting bigger windows, and it won’t matter: unmetered budgets always fill. Teams that treat tokens like compute will ship faster agents at a fraction of the cost, and the gap will look like magic from outside. It’s just profiling.
