The coordination limit
Researchers at the University of Konstanz studied how language models behave when many agents must repeatedly choose between equally valid options. In small groups, agreement arrived quickly. As the group grew, consensus took longer and the risk of splitting into factions increased.
The most capable systems were able to maintain agreement across groups approaching a thousand agents, while simpler models coordinated effectively with much smaller groups.
Consensus can amplify mistakes
The same mechanism that helps a group converge can also suppress independent judgment. If each agent automatically follows the majority, a group may become highly confident in a decision that no individual would have selected alone.
Why this matters for security
Multi-agent systems are increasingly being considered for monitoring, triage, automation, and operational decisions. A reliable design must preserve dissent, record the reasoning behind important decisions, and require human review when the consequences are significant.
Agreement is useful evidence. It is not proof.
← Back to FXeSEC