37%
higher quality than agent exploration
Same question, same project, two windows: graph-first answered “how does this AI pipeline work” in 2m3s for $1.02 and scored 24 of 25 — against Explore agents’ 5m52s, ~$2.50, and 17.5 of 25. The graph cited 8+ files with line counts where agents cited one.
The graph queried structural relationships, then read the exact files it surfaced — and found what broad exploration missed: a 9-step pipeline (not a tidy 4-phase model), safety guardrails running first at step 0, first-class cost governance, and a staged-write cache. The Explore agent returned a plausible summary that flattened all of it.
The lesson the test corrected its own hypothesis on: agents aren’t the depth layer — reading is. The graph tells you where; reading the files it found tells you what and why, in the same context, with no agent spawn. Agents are for parallel writes and overflow, not for understanding code.
Not features for their own sake — measurable leverage on every session.