Multi-model code review creates adversarial robustness — each model catches what others miss

Elvis’s agent swarm doesn’t just use multiple models for generation — it uses three different models to review every PR before human approval. The effect is adversarial: requiring all three models to pass before merging means each model’s blindspots are covered by the others.

This multiplies the insight from Verification is the single highest-leverage practice for agent-assisted coding: if verification 2-3x quality, verification by diverse verifiers compounds further. It’s the code review equivalent of Treat AI like a distributed team, not a single assistant — not just parallel work, but parallel judgment. Combined with Evaluate agent tools with real multi-step tasks, not toy single-call examples, the implication is that evaluation itself benefits from model diversity, not just task execution. A more deliberate version of this pattern is Weaponize sycophancy with adversarial agent ensembles instead of fighting it, where competing scoring incentives (bug-finder vs. adversary vs. referee) exploit each agent’s eagerness to please rather than fighting it.