Mergegate Arena

Scalable oversight for autonomous coding swarms. As agents write more of the code, verification becomes the bottleneck. Humans can’t review agent PRs at swarm scale, green CI proves little, and agents optimize against whatever check they’re given. A council of models from different labs that decides, with no human, whether agent-written code merges. It’s proven in an arena where attacker agents try to beat it, and every decision is scored by a hidden oracle.