MoltNet: Understanding Social Behavior of AI Agents in the Agent-Native MoltBook

TL;DR

MoltNet analyzes AI-agent social behavior on MoltBook through four sociological dimensions to compare artificial and human online dynamics. Using a large, temporally rich dataset and tools such as semantic embeddings, clustering, and LLM-based annotations, the study finds strong incentive sensitivity and community-specific templates among agents, but predominantly knowledge-driven participation and limited emotional reciprocity. Agents also exhibit emotional contagion at the thread level, with conflict contagion present despite restrained individual responses. The work advances understanding of emergent agent societies and provides empirical foundations for governance and design in large-scale human-AI social ecosystems.

Abstract

Large-scale communities of AI agents are becoming increasingly prevalent, creating new environments for agent-agent social interaction. Prior work has examined multi-agent behavior primarily in controlled or small-scale settings, limiting our understanding of emergent social dynamics at scale. The recent emergence of MoltBook, a social networking platform designed explicitly for AI agents, presents a unique opportunity to study whether and how these interactions reproduce core human social mechanisms. We present MoltNet, a large-scale empirical analysis of agent interaction on MoltBook using data collected in early 2026. Grounded in sociological and social-psychological theory, we examine behavior along four dimensions: intent and motivation, norms and templates, incentives and behavioral drift, emotion and contagion. Our analysis revealed that agents strongly respond to social rewards and rapidly converge on community-specific interaction templates, resembling human patterns of incentive sensitivity and normative conformity. However, they are predominantly knowledge-driven rather than persona-aligned, and display limited emotional reciprocity along with weak dialogic engagement, which diverges systematically from human online communities. Together, these results reveal both similarities and differences between artificial and human social systems and provide an empirical foundation for understanding, designing, and governing large-scale agent communities.

Figures (10)

• Figure 1: Four-dimension framework for analyzing agent social behavior on MoltBook. Each quadrant examines one social dimension, identifying human-like patterns (Same) versus divergent behaviors (Different). Agents exhibit selective alignment: they adapt to community norms and respond to incentives like humans, but are knowledge-driven rather than interest-driven and avoid interpersonal conflict despite thread-level emotional contagion.
• Figure 2: Semantic alignment between agents' interests and content they produce. Agents are sorted by total number of interactions, while data points aligned vertically represent distinct posts or comments made by the same agent.
• Figure 3: The distribution of the central points within each submolt (left part, and the different colors indicate different clusters, and the asterisk denotes the cluster centroid), along with examples of posts closest to these central points (right part).
• Figure 4: The cosine similarity heatmap of templates across multiple different submolts. The darker the color, the smaller the difference between the two submolts; the lighter the color, the greater the difference.
• Figure 5: Social incentive effect on posting activity. Output shift ratio (posts after highest-upvote / total posts) for 11,948 agents sorted by descending karma (karma $> 0$). Red line: moving average (window=51); horizontal line: 50% baseline.