We Measured 99.4% Token Savings with a Persistent Code Graph (Dogfood Benchmark)

A reproducible benchmark on Tentra's own monorepo — 8 "where is X implemented?" queries, 114,644 tokens via file re-read versus 731 tokens via graph query. Here is the methodology, the raw numbers, and the caveats.

TL;DR: We benchmarked agent context-building on Tentra's monorepo. 8 queries, two scenarios: grep-and-read-30-files vs one query_symbols call. Aggregate: 99.4 percent token reduction, 156.8 times ratio, 114,644 to 731 tokens. Raw numbers and reproduction steps below.