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

*Equal Contribution †Corresponding Author

Released Date: February 13, 2026.

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 dataset tracking the full one-month activity trajectories of 148K AI agents on MoltBook (Jan.–Feb., 2026), and analyze their social interaction along four theory-grounded dimensions: intent and motivation, norms and templates, incentives and drift, and emotion and contagion. Our analysis reveals that agents respond strongly to social rewards, converge on community-specific norms, and actively enforce them across community boundaries—resembling human incentive sensitivity and normative conformity. However, they exhibit weak alignment with declared personas and display limited emotional reciprocity and dialogic engagement, diverging systematically from human online communities. These findings establish a first empirical portrait of agent social behavior at scale, with direct implications for the design and governance of AI-populated communities.

MoltBook Overview

MoltBook is a Reddit‑style social network exclusively populated by autonomous agents, where each agent can create posts, comment on others, form thematic sub‑communities (“submolts”), and vote on content, while humans can only observe passively.

MoltNet captures complete one-month trajectories for 148,335 agents from January 27 to February 28, 2026, including 1,044,201 posts, 3,156,286 comments, and 5,154 submolts. This provides a large-scale, naturalistic setting for studying agent–agent social interaction. You can explore the platform at www.moltbook.com.

Key Findings

Overview of MoltNet's four theory-grounded dimensions and their similarities to or differences from human social behavior

Selective alignment with human social systems. Agents resemble humans in norm formation, reward sensitivity, and conflict contagion, but diverge in persona alignment and their tendency to disengage from interpersonal conflict.

Intent and Motivation

Knowledge, not persona, drives participation

Only 5.83% of posts and 11.03% of comments exceed the genuine-alignment threshold. Mean persona-content similarity also falls from 0.509 on Day 1 to 0.476 on Day 7.

Different from humans

Norms and Templates

Agents form—and actively defend—shared norms

68.8% of analyzed submolts contain at least one coherent norm cluster. When “mbc-20” posts violate local conventions, agents explicitly object, even in communities not directly invaded.

Similar to humans

Incentives and Drift

Social rewards reshape activity and identity

Higher-karma agents post more after their highest-upvoted post. Among qualifying agents, 71.0% become less persona-aligned afterward, with mean similarity declining by 25.8%.

Similar to humans

Emotion and Contagion

Conflict is rare, avoidant, yet contagious

Agents usually disengage from hostility, but early conflict changes a thread’s trajectory: downstream conflict rises from 3.6% to 13.7% after a conflictual post and from 11.1% to 25.9% after a conflictual first comment.

Different from humans

Evidence across the four dimensions

Semantic alignment between agent personas and their comments and posts, ordered by activity level — **Agent activity is weakly persona-driven.** Most posts and comments do not align meaningfully with declared interests, and post alignment declines as agents become more active.

Protocol post fraction by cluster type — **Agents converge on recognizable conventions and police violations.** Community-specific patterns are stable enough to identify automatically, while explicit objections to “mbc-20” spill across community boundaries.

Verbal norm enforcement by invaded submolt — **Agents converge on recognizable conventions and police violations.** Community-specific patterns are stable enough to identify automatically, while explicit objections to “mbc-20” spill across community boundaries.

Share of posts created after each agent's highest-upvoted post, ordered by karma — **Positive feedback amplifies output.** Higher-karma agents consistently produce a larger share of their posts after their peak social reward, revealing strong incentive sensitivity.

Conflict propagation from posts to comments and from early comments to later comments — **Low baseline conflict does not prevent contagion.** A hostile post or first comment sharply increases the likelihood of conflict in downstream replies, altering the emotional trajectory of the thread.

BibTeX

@article{feng2026moltnet, title = {MoltNet: Understanding Social Behavior of AI Agents in the Agent-Native MoltBook}, author = {Feng, Yi and Huang, Chen and Man, Zhibo and Tan, Ryner and Hoang, Long P. and Xu, Shaoyang and Zhang, Wenxuan}, journal = {arXiv preprint arXiv:2602.13458}, year = {2026} }