30×
fewer tokens than agent exploration
Code-graph queries replace agent fan-outs
Five targeted graph queries found the same files three parallel Explore agents did — about 500 words instead of 9,000, in under five seconds instead of forty-five. The graph already knows the structure; agents rediscover it every time.
Mapping a feature's files the usual way means fanning out parallel exploration agents — each one starts with empty context, reads broadly, and reports back a summary. That run cost about 9,000 words and 45+ seconds.
Five targeted code-graph queries returned the same files in roughly 500 words, in under five seconds, with no agent spawns at all. Structure that's already indexed doesn't need rediscovering — the graph answers callers, callees, and blast radius instantly.
- 5 graph queries replaced 3 Explore agents
- ~500 words vs ~9,000
- <5s vs 45s+, with no cold-context spawns
This is what Pro delivers.
Not features for their own sake — measurable leverage on every session.