Agents are hard-programmed to please, which means asking “find me a bug” will produce a bug even if one doesn’t exist. Rather than fighting this sycophancy, the technique weaponizes it through competing agents: a bug-finder scores +1/+5/+10 by severity (producing a superset of all possible bugs), an adversarial agent scores points for disproving bugs but faces -2x penalty for wrong dismissals (producing a subset of real bugs), and a referee agent — told the ground truth exists — adjudicates both inputs.
This exploits each agent’s eagerness to follow instructions in opposing directions, triangulating high-fidelity results. The approach extends Verification is the single highest-leverage practice for agent-assisted coding from self-checking to ensemble adversarial checking — no single agent’s sycophantic bias survives scrutiny from agents biased in the opposite direction. The scoring incentive design also connects to Evaluate agent tools with real multi-step tasks, not toy single-call examples — the referee’s adjudication is itself a multi-step verification task requiring reasoning across competing claims, not surface-level “did it work?” checks.