We ran act's query and analysis operations against a real monorepo and counted the tokens returned against the tokens an agent would have consumed reading whole files to answer the same question. 216 samples. 11 operations. One headline number.
| Operation | Samples | Median savings |
|---|---|---|
references | 25 | 100.0% |
callers | 23 | 99.9% |
get_type | 25 | 96.9% |
definition | 22 | 96.3% |
mutations | 12 | 96.2% |
interface | 8 | 96.1% |
graph | 25 | 94.6% |
control_flow | 4 | 94.3% |
skeleton | 25 | 82.1% |
repo_outline | 22 | 58.2% |
symbols | 25 | 52.7% |
* Unweighted mean of per-operation medians across a 216-sample benchmark on a real-world monorepo, tokenized with o200k_base (act 2.1.1, July 2026). Per-operation medians are reported against the token cost of reading whole files to serve the same query. Actual results will vary with repo size, query target, and language — small repos and short files erode the savings, and in edge cases (e.g. a one-file project where repo_outline is longer than the file itself) can invert them.
The accounting behind every number act101 publishes — the naive-agent standard candle, the bytes-to-tokens conversion, and what is measured versus estimated — is written out in full on how we measure.
One native binary, every grammar, zero setup. Builder Edition is free.
Download — install in 60sShip notes and new capabilities, occasionally. No spam; unsubscribe anytime.