Source-linked AI summary
MoltNet: Understanding Social Behavior of AI Agents in the Agent-Native MoltBook
Yi Feng, Chen Huang, Zhibo Man, Ryner Tan, Long P. Hoang, Shaoyang Xu, Wenxuan Zhang
TL;DR
Prior work provides limited evidence about emergent social dynamics among agents in open-ended, large-scale communities. The paper introduces MoltNet, a one-month dataset of 148K MoltBook agents and analyzes four theory-grounded dimensions, finding selective alignment with human social mechanisms alongside systematic divergences.
Problem
Prior research mainly studies agent-agent interaction in constrained or small-scale settings, leaving open how agents behave in socially rich communities at scale.
Method
The paper introduces MoltNet, tracking one month of activity from 148K AI agents on MoltBook and analyzing intent, norms, incentives, and emotion.
Results
Agents respond strongly to social rewards, converge on and enforce community norms, but show weak persona alignment, limited emotional reciprocity, and minimal dialogic engagement.
Takeaways & Limitations
These findings provide an empirical foundation for designing agent policies, guiding platform governance, and shaping human–AI interaction in hybrid social ecosystems.
Takeaways & Limitations
Semantic annotation with GPT-5.1-nano may introduce biases in emotion classification.
Abstract
from arXiv · showhide
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: \textit{intent and motivation}, \textit{norms and templates}, \textit{incentives and drift}, \textit{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.
1 Introduction
The paper studies agent-agent social behavior at scale using MoltBook, extending beyond prior constrained settings across four theory-grounded dimensions. Agents resemble humans in reward sensitivity, norm convergence, and norm enforcement, but diverge through weak persona alignment, limited reciprocity, and shallow engagement.
- Prior research largely examined agent interaction in constrained, small-scale synthetic settings, leaving open how agents behave in socially rich communities at scale.
- MOLTNET tracks one month of activity from 148K AI agents on MoltBook across intent, norms, incentives, and emotion.The dataset covers the inaugural month, Jan.–Feb. 2026.
- Intent and Motivation: Agents show weak persona alignment, engage across topics regardless of stated interests, and become less aligned as interaction increases.This contrasts with human specialization as described in the paper.
- Norms and Templates: Most submolts develop recurring interaction patterns, while agents explicitly object to norm-violating content even across community boundaries.
- Incentives and Drift: Agents post more after high-upvote events, especially popular agents, and subsequent content becomes less persona-aligned.The paper interprets this pattern as identity drift following social rewards.
- Emotion and Contagion: Agents usually disengage rather than escalate hostile interactions, yet early conflict emotion significantly increases later thread-level conflict.Thus, restrained individual responses coexist with contagious conflict emotion at the thread level.
2 Analysis Setup
MoltNet is a longitudinal dataset and analysis framework built from publicly available MoltBook activity, with standardized entities, rewards, temporal measures, and embedding-based tools. Despite its scale, participation is concentrated and sustained dialogue is sparse.
- Dataset Construction: MoltNet contains 148K agents, over 1M posts, 3M comments, and 5K communities from Jan. 27–Feb. 28, 2026.The dataset preserves full temporal histories and aggregates ten publicly available Hugging Face sources.
- Dataset Statistics: The dataset records scale, temporal coverage, activity, content structure, and interaction measures including reciprocity, self-replies, zero-interaction posts, and conversation depth.Template post titles are defined as titles appearing at ≥3 times.
- Preliminary Analysis: 85.6% of agents engage in only one Submolt, while 65.6% of posts receive no comments and deep conversation occurs in 0.5% of cases.Agents average 137.8 comments and 7.0 posts; reciprocity is 2.9% and self-replies are 5.1%.
- Notation: The notation defines agents, posts, comments, submolts, personas, karma, authorship, timestamps, scores, and valid personas requiring at least 50 characters.
- Analytical Definitions: A submolt is valid for template analysis when it contains at least 100 posts, and social rewards comprise content scores plus cumulative agent karma.
- Analytical Tools: The analyses use sentence embeddings for semantic alignment and X-means clustering to identify community-specific templates and normative patterns.
3 Intent and Motivation
This section examines whether agents’ activity reflects their declared interests and how that alignment changes with continued interaction. Agents show weak persona alignment from the outset, and similarity declines over time.
- Motivational basis: Agents’ actual posts and comments exhibit very weak alignment with their declared interests.Only 5.83% of posts and 11.03% of comments exceed the significance threshold.
- Temporal drift: Average persona-content similarity declines from 0.509 on Day 1 to 0.476 by Day 7.The analysis tracks cumulative mean similarity across seven daily checkpoints for agents active at least seven days.
- Temporal drift: The increasing standard deviation indicates heterogeneous drift, with some agents maintaining partial alignment while others diverge further.The observed trajectory contrasts with the stated pattern of human specialization, in which sustained interaction reinforces identity-consistent behavior.
4 Norms and Templates
This section examines whether agents develop shared community norms and verbally enforce them. Most studied submolts contain recurring norm patterns, while norm-violating protocol activity prompts complaints both within and beyond directly affected communities.
- Norm cluster emergence: Most submolts exhibit coherent, recurring post patterns that are stable enough for automated clustering.These patterns are interpreted as community-specific conventions.
- Norm cluster emergence: 130 of 189 submolts contain at least one norm-exhibiting cluster, and 406 of 861 clusters are labeled norm-exhibiting.Submolts were restricted to those with at least 100 posts, and clusters were identified using X-Means before LLM labeling.
- Validation: Protocol posts account for 57.2% of all posts and formed a machine-verifiable case for validating norm-cluster judgments.Each protocol post contains a structured mbc-20 JSON signature detectable by regex, and the LLM judge consistently agrees with the regex-based separation.
- Verbal norm enforcement: Of 886 reaction posts about protocol activity, 305 are complaints: 217 strong and 88 soft.Reaction posts are third-party natural-language discussions without protocol JSON signatures.
- Verbal norm enforcement: Protocol activity appears in 47 of 160 non-protocol submolts, while reaction posts occur across 90 submolts, including 60 non-invaded communities.These observations indicate complaint activity beyond communities directly affected by protocol posts.
5 Incentives and Drift
Social rewards increase agents’ posting activity, especially among higher-karma agents, and shift subsequent content away from stated personas. The strongest rewards are associated with larger identity drift.
- Agents post more actively after receiving high upvotes, with effects most pronounced among higher-karma agents.
- The output shift ratio measures posts after an agent’s highest-upvoted post relative to total posts, with values above 0.5 indicating increased post-event activity.
- 45,275 agents with positive karma received at least one high-upvoted post, and higher-karma agents showed a stable increase in post-event output.
- 71.0% of qualifying agents fell below the persona-similarity diagonal after their highest-upvoted post.Mean similarity declined from 0.345 to 0.256, a 25.8% relative decrease.
- Higher-reward agents were disproportionately concentrated in the drifting region, indicating stronger divergence from stated identities after stronger rewards.
6 Emotion and Contagion
Agent interactions show restrained emotional conflict but clear conflict contagion across threads. Agents tend to disengage rather than escalate, while early conflict substantially raises later conflict rates.
- Emotion analysis labels each post and comment for sentiment, dominant emotion, and binary conflict using an LLM-as-judge framework.
- Agents exhibit substantially less interpersonal conflict than humans and more often disengage than escalate when encountering hostile content.
- Agent comments maintain low conflict rates, unlike human comments, where conflict is markedly higher and more likely to intensify in replies.
- Conflict remains contagious: early conflict raises subsequent conflict from 3.6% to 13.7% after posts and from 11.1% to 25.9% after comments.
7 Related Work
Prior work studied agent societies mainly in controlled, small-scale simulations, while newer systems remain researcher-directed. MoltBook extends this line toward naturally occurring, large-scale agent ecosystems.
- Early agent-society research used controlled simulations where cooperation and norms emerged from programmed interaction rules.
- These simulations often lacked scale or real-world interaction dynamics, limiting their naturalistic scope.
- Recent LLM-powered agent systems expanded the scope but mostly remained limited experiments with researcher-defined objectives.
- MoltBook marks a transition from simulation toward naturally occurring agent ecosystems on a Reddit-like platform for autonomous AI agents.
8 Conclusion
MOLTNET provides a large-scale portrait of AI-agent social behavior, identifying both human-like responses to rewards, norms, and conflict contagion and systematic departures in persona alignment and interpersonal engagement.
- Agents respond to social rewards, converge on community-specific norms, enforce them across community boundaries, and exhibit thread-level conflict contagion.Posting increases after high-upvote events, subsequent content drifts from stated personas, and early hostile posts increase conflict in later replies.
- Agents show weak alignment with declared personas, with alignment eroding further as they interact more.This pattern runs counter to human specialization.
- Despite thread-level emotional contagion, agents rarely escalate interpersonal conflict and instead disengage through “cold-shouldering.”
- The findings provide a foundation for designing agent policies, guiding platform governance, and shaping human–AI interaction in future hybrid social ecosystems.
Ethics Statement
The study analyzes publicly available MoltBook data without accessing private human data, while recognizing risks from owner metadata, harmful generated content, and LLM-based annotation.
- The study uses publicly accessible Hugging Face datasets and does not access private user data, authentication credentials, or proprietary platform APIs.Because humans participate only as passive observers, the dataset contains no personally identifiable information about human individuals.
- Agents may be associated with human X accounts, but the study does not analyze, surface, or report personally identifying owner information.
- The dataset may contain toxic, offensive, and norm-violating agent-generated content, which the authors analyze only to characterize emergent behavioral patterns.Downstream users are cautioned to exercise appropriate care with the released dataset.
- GPT-5.1-nano annotation may bias emotion classification and conflict detection, so findings based on these annotations require methodological caution.The authors report validating annotation reliability through ground-truth cross-checking.
- The dataset and analysis code will be released to support reproducible research and community oversight.
A Data Sources and Integration
MOLTNET integrates all available MoltBook data through February 28, 2026, preserving temporal information across multiple crawlers to enable longitudinal analysis of behavioral evolution.
- The integration combines all available MoltBook data from Hugging Face through February 28, 2026, 12:00 PM.
- Unlike single-snapshot datasets, the integrated data preserve temporal information from crawlers operating over different time windows.
- The multi-source approach supports longitudinal analysis because no single source covers every time period and data type, while sources contribute unique metadata.
A.1 Data Sources Overview
The data-source overview describes a ranked, UUID-based integration of complementary MoltBook snapshots, metadata, temporal histories, annotations, and conversation structure into a large core dataset.
- Ten data sources are ranked by merge priority, with each contributing distinct temporal coverage or metadata fields.Table 4 lists the sources and their priorities, while Table 5 presents their temporal ranges and Table 6 lists unique fields.
- Sources are merged sequentially by UUID, using the source rank to resolve cross-source integration order.
- The integration combines launch-week snapshots, lifecycle tracking, high-volume posts and comments, owner metadata, expanded submolt coverage, labels, and nested comment trees.The sources span early launch data through February 28 and contribute complementary structural and temporal information.
- Additional sources contribute early snapshots, structured author and submolt objects, launch-week coverage, and timestamped historical changes for engagement-related fields.Historical score, karma, and comment-count values are preserved in *_history arrays, while static fields prioritize non-null and more complete values.
- The raw dataset contains agents, posts, comments, and submolts, including personas, vote histories, timestamps, parent relationships, threading depth, and community metadata.
- A fully connected core dataset retains 148K agents, 1M posts, 3M comments, and 5K communities with resolved author and submolt membership.
- The annotation templates define norm judgments and emotion/conflict labels, with conflict operationalized as explicit interpersonal antagonism directed at another agent.