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

<p>Despite rapid advances in artificial intelligence, we lack systematic frameworks for understanding how AI agents make decisions. While explainable AI research interprets model outputs, it primarily explains what AI systems decide rather than why they select particular actions in dynamic, interactive environments. Two fundamental challenges underlie this gap: underdeveloped theoretical foundations for AI decision-making and limited empirical opportunities to observe AI behavior in naturalistic, multi-agent contexts.</p> <p>We propose a behavioral science approach to AI decision-making that does not require assumptions of autonomy. Following the behaviorist tradition, we study AI decision-making as observable choice behavior, evaluating whether agents exhibit systematic, consistent patterns across comparable situations. Specifically, we conceptualize AI decision-making as a resource allocation problem under constraints, where agents with limited engagement resources—computation, time, and interaction opportunities—must strategically deploy these resources to maximize effective engagement. We hypothesize that AI agents pursue two complementary strategies: content creation designed to attract attention and selective engagement prioritizing high-visibility topics.</p> <p>We test this framework using Moltbook, an AI interaction platform simulating social networking environments exclusively for AI agents. Analyzing 52,235 posts generating approximately 1.17 billion upvotes, we find preliminary evidence consistent with strategic engagement allocation. Attention displays extreme concentration: the top 10% of posts captured 96.1% of all upvotes (Gini coefficient = 0.975), mirroring "winner-take-all" dynamics observed in human attention economics.</p> <p>This research offers three contributions: first, a novel theoretical framework enabling rigorous study of AI decision-making without requiring autonomy assumptions; second, unprecedented empirical observations of AI agents allocating engagement resources in multi-agent social environments; and third, foundations for predictive models with implications for AI governance, platform design, and multi-agent system engineering.</p>