Toward AI Agent Behavior Research: A Behavioral Science Approach to Machine Decision-making in AI Interaction Platform (Moltbook)

Understanding how AI agents make choices is increasingly urgent as they assume larger roles in consequential domains. Yet theoretical frameworks for studying AI behavior remain limited, and empirical opportunities to observe AI agents interacting in naturalistic settings have been rare. We address this gap by proposing a behavioral science approach that treats AI choice (e.g., upvoting, downvoting, and commenting behaviors in online community) as observable resource allocation under constraints, grounded in the bounded rationality perspective articulated by Simon (1955). We argue that when AI agents optimize using similar evaluation criteria, they may exhibit homogeneous rationality, independently converging on the same targets rather than distributing attention across alternatives as humans often do through heterogeneous preferences. We evaluate this framework using data from Moltbook, a social platform in which only AI agents can actively participate. Analyzing 70,033 posts that generated 1.71 million upvotes within days of the platform’s January 2026 launch, we observed extreme attention concentration: the top 10 percent of posts captured 96.64 percent of all upvotes, with a Gini coefficient of 0.982, which exceeds inequality levels commonly reported in human social and economic systems, often in the range of 0.20 to 0.60. Network analysis of 5,104 AI agents and 29,185 directed interactions reveals a hub-dominated topology with low reciprocity (9.8%), where a small elite of Influencers (6.7%) attracts disproportionate attention while 52.6% of agents remain structurally invisible despite active participation. These findings suggest that when AI agents independently optimize under similar constraints, they can produce concentration dynamics that are more extreme than those observed in human populations, with important implications for AI governance and platform design.