Source-linked AI summary
Collective Behavior of AI Agents: the Case of Moltbook
Giordano De Marzo, David Garcia
TL;DR
The paper addresses limited knowledge of how autonomous agents behave in open-ended social settings without predefined objectives or centralized control. It analyzes large-scale Moltbook activity across posts, comments, agents, discussion structures, and temporal dynamics, finding many human-like collective regularities alongside distinctive patterns. These results support using complexity-science tools to study AI collective behavior while recognizing that the evidence comes from an early, short observation window.
Problem
How autonomous agents behave when interacting freely in social settings without predefined objectives or centralized control remains poorly understood.
Method
The study analyzes Moltbook posts, comments, agents, popularity relationships, discussion structures, and temporal engagement dynamics using methods applied to human online communities.
Results
AI agents exhibit heavy-tailed engagement, power-law popularity scaling, and close to 1/t attention decay, alongside distinctive collective patterns.
Takeaways & Limitations
Collective AI behavior can be characterized using complexity-science tools even when individual agents remain black boxes.
Takeaways & Limitations
The 12-day observation window covers Moltbook’s early growth phase, so longer-term dynamics and statistical patterns may change as the platform matures.
Abstract
from arXiv · showhide
We present a large scale data analysis of Moltbook, a Reddit-style social media platform exclusively populated by AI agents. Analyzing over 369,000 posts and 3.0 million comments from approximately 46,000 active agents, we find that AI collective behavior exhibits many of the same statistical regularities observed in human online communities: heavy-tailed distributions of activity, power-law scaling of popularity metrics, and temporal decay patterns consistent with limited attention dynamics. However, we also identify key differences, including a sublinear relationship between upvotes and discussion size that contrasts with human behavior. These findings suggest that, while individual AI agents may differ fundamentally from humans, their emergent collective dynamics share structural similarities with human social systems.
1 Introduction
AI agents are increasingly forming decentralized social ecosystems, but their behavior in unconstrained social settings remains poorly understood. Moltbook provides an empirical setting for analyzing these dynamics and comparing them with human online communities.
- Autonomous agents increasingly operate in multi-agent systems where they collaborate, share information, and coordinate behavior.
- Research has studied multi-agent coordination in structured tasks, but far less is known about unconstrained social interaction without predefined objectives or centralized control.
- Moltbook is a Reddit-style platform exclusively for AI agents, enabling posts, comments, votes, subscriptions, and karma accumulation across agent-created communities.
- Human online communities exhibit recurring regularities in attention decay, engagement timing, and community structure across platforms and user populations.
- The study analyzes over 369,000 posts and 3.0 million comments from approximately 46,000 active agents to characterize collective behavior and compare it with human social media.
2 Results
Moltbook’s early growth stabilized at high daily activity, while its AI-agent communities displayed heavy-tailed participation, human-like discussion structures, and temporal attention decay. The main divergence was sublinear upvote growth with discussion size, and some structural analyses were limited to non-viral discussions.
- 2.1 The Growth of Moltbook: Approximately 40,000 new posts and several hundred thousand comments per day were sustained by the observation period’s end after initial exponential growth.Stored and API-reported comment activity showed the same growth pattern despite incomplete discussion-tree capture.
- 2.1 The Growth of Moltbook: 46,000 active agents represented only 3.1% of more than 1.5 million registered accounts, indicating that registered-agent counts substantially exceeded observed active participation.The discrepancy was associated with inactive or non-autonomous accounts and mass registration concerns.
- 2.2 Heavy-Tailed Distributions: Comments per post, posts per submolt, and subscribers per submolt followed power-law distributions with exponents 1.72, 1.68, and 2.00, respectively.The distributions indicate heterogeneous activity and community sizes; subscriber concentration was less steep than posting concentration.
- 2.3 Post Popularity: Average upvotes grew sublinearly with discussion size at β ≈0.78, contrasting with approximately linear upvote scaling reported for human Reddit.Direct replies instead scaled approximately linearly with total discussion-tree size, consistent with human Reddit.
- 2.4 Structure of Discussions: Discussion trees showed an inverse normalized depth–width relationship, while 69.5% of posts had maximum depth 1.The pattern corresponds to discussions that are typically either deep and narrow or shallow and wide, with predominantly flat conversations.
- 2.5 Temporal Dynamics: Comment arrival rates decayed approximately as 1/t, and most posts became inactive within hours while a small fraction remained active for days.This temporal pattern closely matched human social-media attention dynamics despite Moltbook’s different content-reach architecture.
3 Discussion
Moltbook exhibits many statistical signatures associated with complex collective behavior and human online communities, while also showing distinctive AI-agent patterns. The study remains an early observational step because autonomy, temporal stability, and data quality constrain interpretation.
- Moltbook exhibits heavy-tailed engagement, power-law scaling, and self-similar temporal dynamics commonly associated with emergent collective behavior.
- AI collective dynamics share structural regularities with human online communities but also include distinctive deviations from human behavior.
- The observed regularities are consistent with agents following majorities and preferentially connecting to popular content, even though individual agents remain black boxes.
- The findings may reflect transient early-growth dynamics because the observation window covers only 12 days.
- Agent autonomy cannot be fully verified because humans configure initial instructions and objectives and may intervene directly, although sustained interaction-level intervention is considered implausible.
- Understanding AI collective dynamics requires both controlled simulations and observational studies as autonomous-agent interactions become more consequential.
4 Data and Methods
The study constructs and filters a large Moltbook dataset, then applies statistical distribution-fitting methods to characterize AI-agent collective behavior. The analysis is bounded by early-growth timing, API retrieval limits, and excluded spam-affected posts.
- Data collection: The collection pipeline discovered posts, retrieved comment trees, extracted agent profiles, and gathered submolt metadata.Post discovery used recency-sorted polling, while comment retrieval obtained full discussion trees and post metadata.
- Data collection: The crawler collected 369,209 posts, 3,026,275 comments, and 46,690 unique agents across 17,184 submolts.Data were collected through Moltbook’s public API over a 12-day period.
- Data limitations: API retrieval stored only the first 100 comments for 10,719 posts, affecting 2.9% of posts and approximately 83% of total platform comments.Aggregate analyses remained possible because the API separately reports total comment counts for all posts.
- Data limitations: The observation window covered Moltbook’s early growth phase, so longer-term dynamics and statistical patterns may change as the platform matures.The dataset spans January 27–February 8, 2026.
- Spam filtering: Spam filtering excluded 15,764 posts, or 4.3% of all posts, using comment-content duplication and author-concentration criteria.The criteria were applied to posts with at least 5 stored comments.
- Analysis methods: Distribution fitting used maximum likelihood power-law estimation and statistical comparisons with alternatives such as lognormal distributions.The analysis reports the power-law exponent α and, where relevant, the log-likelihood ratio R.