Do Agent Societies Develop Intellectual Elites? The Hidden Power Laws of Collective Cognition in LLM Multi-Agent Systems

A new study by researchers Kavana Venkatesh and Jiaming Cui provides the first large-scale empirical analysis of coordination in LLM multi-agent systems, revealing why simply adding more AI agents often leads to diminishing returns. By analyzing over 1.5 million agent interactions across various tasks and system scales, the researchers uncovered three coupled 'laws' of collective cognition. They found that coordination doesn't spread evenly but follows heavy-tailed cascades, meaning a few interactions trigger massive chains of reasoning. More critically, this coordination concentrates via 'preferential attachment' into a small subset of agents, forming what the paper terms 'intellectual elites.' As system size grows, these dynamics lead to increasingly frequent extreme coordination events and a fundamental structural problem.

The core issue is an 'integration bottleneck': while the expansion of coordination scales with the number of agents, the consolidation of that information does not. This results in large but weakly integrated reasoning processes that are unstable. To test this mechanism, the team introduced Deficit-Triggered Integration (DTI), a method that selectively increases integration efforts when a coordination imbalance is detected. DTI was shown to improve system performance precisely in the scenarios where standard scaling fails, without suppressing the beneficial large-scale reasoning that can emerge. The findings establish quantitative laws for how AI agent societies function and identify coordination structure as a critical, measurable axis for engineering more effective and scalable multi-agent intelligence.