One third the tokens, for the same change
Early internal testing on an estate-wide change, one like-for-like comparison. Why scoped context beats a bigger context window.
A generic agent spends its budget discovering what a graph already knows: which services a change touches, who owns them, what breaks downstream. It greps, guesses and re-reads, once per repository.
The measurement
On an estate-wide change spanning multiple repositories, the case where a generic agent's discovery cost is worst, the same outcome cost roughly a third of the tokens when each lane received a scoped context pack instead of an estate to explore.
This is early internal testing and one like-for-like comparison, not a published benchmark. The direction is what matters more than the number.
Why a bigger window does not fix it
Context windows grow and the bill grows with them. The problem is not how much an agent can hold, it is how much it has to read to find out what matters. A lookup does not scale with the size of the estate. Exploration does.
The larger and messier the customer, the wider that gap opens, which is exactly the customer worth having.