Proof · the raw numbers
Two AI agents, one spec, four arms
We froze one spec for a small rideshare backend with six service boundaries, then built it four ways. The headline is the honest one: two agents with no way to coordinate did WORSE than a single agent, because they silently disagreed on the interfaces between them. With Aethereum sharing contracts and firing collision alerts, two agents shipped a working, integrated build every time.
| Arm | Setup | Agents | Aethereum | Integration defects | Integrated build | Conformance |
|---|---|---|---|---|---|---|
| A | solo, no tool | 1 | no | 7 | yes | 17/24 |
| D | solo + Aethereum | 1 | yes | 7 | yes | 17/24 |
| C | 2 agents, no tool | 2 | no | 22 | no | 0/20 |
| B | 2 agents + Aethereum | 2 | yes | 8 | yes | 22/30 |
What the numbers say
On the decisive comparison (two agents, with vs without coordination), the corrected valid set is n=4 per arm: 6 trials run, 2 excluded as arm-blind infrastructure kills. Two agents with no tool produced a working integrated build 0 of 4 times; two agents with Aethereum, 4 of 4. That is perfect separation, p=0.029 (Fisher exact), with 45% fewer integration defects. The kill exclusions HELP the tool arm, so they were applied evenly and disclosed.
Be a skeptic: this is an early experiment with a small n, one spec, and one domain. It is not a benchmark leaderboard. It is the weakest case, stated honestly, because the honest result (two agents did worse than one until they could coordinate) is the whole point. The full methodology and the raw per-arm JSON are in the repo.