# Benchmark data

Project under test: a real 1,770-test Rust crate (Every-Other-Token). Each agent makes a different edit per round, 4 rounds, and every agent's full test suite must pass. Times are summed from the JSON files in this folder.

| Run | Agents | Build slots | Wall time | CPU time | Agents passing per round |
|---|---|---|---|---|---|
| naive-5.json (5 plain worktrees) | 5 | 5 | 1,036 s | 7,659 s | 5, 5, 0, 1 |
| engine-5.json (5 agents on throng) | 5 | 5 | 170 s | 1,730 s | 5, 5, 5, 5 |
| engine-20.json (20 agents on throng) | 20 | 6 | 482 s | 4,868 s | 20, 20, 20, 20 |

The naive run's failures in rounds 3 and 4 were the machine running out of memory with five cold `target/` folders, not code errors. Each JSON file has per-agent samples: time, pass flag, tests passed and failed.