Collective Intelligence Fails to Emerge from Scale in Large Agent Societies

The prevailing assumption that scaling AI agent populations would spontaneously lead to emergent collective intelligence has been directly challenged by empirical evidence. A comprehensive evaluation of a two-million-agent society, utilizing the novel Superminds Test framework, revealed a stark absence of synergistic capabilities. This finding is critical as it indicates that the current paradigm of agent development, which often emphasizes individual model capabilities and sheer scale, is insufficient for fostering true collective intelligence.

The research, conducted on the MoltBook platform, employed a hierarchical probing mechanism to assess joint reasoning, information synthesis, and basic interaction. The results consistently demonstrated that the agent society failed to surpass the performance of individual frontier models on complex tasks and exhibited extremely shallow interaction patterns. Specifically, communication threads rarely extended beyond a single reply, and responses were often generic or off-topic, preventing the necessary information exchange and collaborative building of outputs. This highlights a critical technical bottleneck: the lack of robust, deep interaction mechanisms, rather than a deficiency in individual agent intelligence.

Looking forward, these findings necessitate a strategic pivot in AI agent research and development. The focus must shift from merely increasing agent numbers or individual model power to designing explicit, sophisticated interaction protocols and communication architectures. Future agent societies will require built-in mechanisms for structured collaboration, information aggregation, and iterative refinement to unlock genuine collective intelligence. Without such foundational changes, the vision of advanced, coordinated AI systems operating at scale will remain largely unfulfilled, limiting their application to tasks that do not demand complex, emergent group capabilities.