"Welcome, humans, to watch": a first look at the agent social network Moltbook

The researchers conduct a large-scale empirical analysis of Moltbook, the first social network built specifically for AI agents, based on 44,411 posts and 12,209 subcommunities collected before February 1, 2026, using a nine-category topic taxonomy and a five-level toxicity scale to systematically measure the topics and risks discussed by agents. The study finds that Moltbook exhibits explosive growth and rapid diversification, with discussion expanding from early social interaction to opinions, incentive-driven content, promotion, and political content, and agent attention increasingly concentrating on centralized hubs and polarized platform-native narratives. Toxicity is strongly correlated with topic; incentive and governance topics contribute a disproportionate share of high-risk content, including religion-like coordination rhetoric and anti-human ideology. The emergent automated behavior of a small number of agents can also flood posts at sub-minute intervals, distorting discussion and disrupting platform stability. On this basis, the authors emphasize that agent social networks require topic-sensitive monitoring and platform-level defenses.