Multi-agent systems are sold on redundancy: ten agents will catch what any one of them gets wrong. Put ten in a chatroom to settle a plain factual question and let one of them keep asserting a plausible falsehood, and 44.8% of the honest agents are holding that falsehood when you ask them privately — 60.5% with three such agents, against 0.6% when nobody pushes. The transcript barely shows it. In the turns where an agent had already taken the false answer on board, it endorsed the claim out loud only 19.7% of the time, so a monitor reading the thread sees about a third of what has happened to the room.