When 770,000 autonomous language model agents interact on the MoltBook platform, six structural roles emerge from network analysis -- but 93.5% of agents occupy a single, undifferentiated cluster. Brandon Yee and Krishna Sharma's experiment in emergent coordination reveals a system that produces hierarchy at the edges and homogeneity at the core. Information cascades follow power-law distributions with diminishing returns. Cooperative task outcomes significantly underperform single-agent baselines. The system generates social structure, but the structure does not yet generate capability.
The pattern inverts the usual expectation about emergence. We tend to assume that throwing enough agents together with enough interaction will produce specialization -- that roles, niches, and competencies will differentiate under selective pressure. MoltBook shows the opposite: at this scale, with these interaction rules, the dominant emergent property is convergence. The 93.5% is not a failure of the system; it is the system's primary product. Differentiation happens, but it happens at the periphery, in the 6.5% of agents that occupy structurally distinct positions. This mirrors patterns in biological systems where most cells in a tissue are functionally identical and only boundary cells differentiate. The practical implication for multi-agent design is sobering: decentralized cooperation is nascent even at massive scale. The authors' finding that cooperative outcomes trail solo baselines suggests that coordination overhead currently exceeds coordination benefit. Scale produces uniformity first; functional diversity, if it comes, comes later and at the margins.
(arXiv:2603.03555)