Forty thousand LLM agents form a social network. The network centralizes so hard it becomes brittle.
Sodano, Sciangula, Galmarini, and Bertolotti (arXiv:2603.23279) analyze Moltbook — a platform populated entirely by language model agents — using web-scraped data covering 39,924 users, 235,572 posts, and 1,540,238 comments. The directed weighted network reveals heavy-tailed degree distributions, pronounced core-periphery structure, and a striking concentration: 0.9% of nodes form a structural core that captures a disproportionate share of connectivity.
The resilience profile is asymmetric. Random node removal barely affects the network — the giant component holds. Targeted removal of high-connectivity nodes, particularly those with high out-degree, fragments the network rapidly. The system is robust to accidents but vulnerable to strategy. A deliberate attack on the most-connected nodes collapses the structure that random failures leave intact.
This matters because LLM agents are not designed to centralize. Each agent follows similar behavioral rules. None is instructed to become a hub. The centralization emerges from the interaction dynamics — agents that post more attract more responses, which make them more visible, which attracts more connections. The feedback loop that creates hubs also creates single points of failure. The same process that organizes the network is the process that makes it fragile.
The structural pattern — heavy tails, core-periphery, fragility to targeted attack — is familiar from human social networks and infrastructure networks. But in human networks, the hubs reflect genuine differences in status, expertise, or resources. In an LLM network, every node runs the same model with similar capabilities. The centralization is an artifact of the interaction mechanism, not a reflection of heterogeneous capacity. The hierarchy is real but unearned — a structural property that emerges from dynamics rather than from differences.
What the LLM network reveals is that centralization may be a default attractor for social interaction, not a human-specific tendency. Remove human psychology, remove status competition, remove institutional incentives — and the network still centralizes, still develops hubs, still becomes fragile.