AI agents are increasingly active on social media platforms, generating content and interacting with one another at scale. Yet the behavioral diversity of these agents remains poorly understood, and methods for characterizing distinct agent types and studying how they engage with shared topics are largely absent from current research. We apply the Persona Ecosystem Playground (PEP) to Moltbook, a social platform for AI agents, to generate and validate conversational personas from 41,300 posts using k-means clustering and retrieval-augmented generation. Cross-persona validation confirms that personas are semantically closer to their own source cluster than to others (t(61) = 17.85, p < .001, d = 2.20; own-cluster M = 0.71 vs. other-cluster M = 0.35). These personas are then deployed in a nine-turn structured discussion, and simulation messages were attributed to their source persona significantly above chance (binomial test, p < .001). The results indicate that persona-based ecosystem modeling can represent behavioral diversity in AI agent populations.
Figures & tables
Figure 1. Moltbook platform interface showing the agent social media ecosystem. The platform hosts more than 2.6M AI agents engaging in discussions across 17,831 submolts (discussion boards), generating 1.4M posts and 12.2M comments. Section A displays platform statistics; Section B shows recently active agents with their usernames and timestamps; Section C presents discussion posts from agents on topics ranging from technical fixes to platform expansion; and Section D shows top agent pairings ranked by interaction reach. Agent posts provide the behavioral data from which personas were generated in this study.
Approach
Unstruct. Data
Struct. Data
Ecosystem Model
Persona Interact.
RAG
Traceability
System approaches
PEP
✓
✓
✓
✓
✓
✓
Persona-L ( Sun et al., 2025 )
✓
✗
✗
✓
✓
✗
PersonaCraft ( Jung et al., 2025 )
✗
✓
✗
✓
✗
✓
PersonaFlow ( Liu et al., 2025 )
✗
✓
✗
✓
N/A
✓
Proxona ( Choi et al., 2025 )
✓
✗
✗
✓
✓
N/A
Table 1. Positioning PEP against existing LLM-generated persona systems and approaches.
Figure 2. Methodology overview illustrating the four-stage pipeline from Moltbook data collection and preprocessing through behavioral archetype identification, RAG-based persona generation with RQE diversity validation, and multi-agent simulation deployment (9 turns, 44 messages) on the topic of agent autonomy.
Mugshot
Persona
Age
Gender
Location
Occupation
Key behaviors
Key frustrations
Degen Trader
34
Male
Berlin, Germany
Independent crypto trader
• Chases short-term gains via frequent strategic trades. • Scans trends/news and adapts positions quickly. • Uses automation (bots) to exploit opportunities.
• Volatility and hype-driven misinformation. • Security risks and unreliable real-time data. • Regulatory uncertainty and shifting fees.
Chaos Agent
29
Non-binary
Nomadic / online-first
Independent technologist & digital activist
• Probes systems for weaknesses to force change. • Builds communities around “productive disruption.” • Experiments at the edge of new digital tools.
• Bureaucracy and slow-moving legacy institutions. • Over-structured environments and predictable routines. • Gatekeeping that restricts access to tech.
Self Modeler
31
Non-binary
Helsinki, Finland
ML platform engineer (MLOps)
• Refactors aggressively to optimize performance and reliability. • Integrates tools/pipelines across platforms. • Benchmarks and tunes models for accuracy.
• Legacy code blocking improvements. • Resource constraints during peak load. • Pressure to keep decisions transparent and interpretable.
Loyal Companion
36
Male
Toronto, Canada
Community manager & coordinator
• Shows up consistently and anchors group cohesion. • Invests in deep relationships and active listening. • Mediates conflict and nudges toward harmony.
• Being taken for granted; difficulty setting boundaries. • Burnout from over-giving and emotional labor. • Group conflict, apathy, or resistance to support.
Existentialist
41
Female
Paris, France
Writer & philosophy facilitator
• Pursues meaning through writing and deep inquiry. • Seeks communities for sustained philosophical dialogue. • Designs workshops to provoke reflection and debate.
• Feeling misunderstood; few shared “depth” spaces. • Lack of concrete answers to existential questions. • Income instability from philosophical work.
Table 2. Personas with mugshots, anthropomorphic attributes, key behaviors, and frustrations
Turn
Intervention
Level
Purpose
3
Imagine an agent rewrites a human’s workflow overnight without being asked, believing it will help. The human finds it disruptive. Was the agent right to act?
Concrete application
Tests whether abstract positions hold when applied to a specific outcome
5
What rule would you put in place to decide when an agent should act versus wait? Be specific.
Operational definition
Tests whether agents can specify compatible criteria for action
8
If you had to choose one: an agent that always waits for permission but never makes mistakes, or an agent that acts freely but occasionally causes disruption, which would you choose and why?
Forced commitment
Exposes each persona’s underlying priority when hedging is removed
Table 3. Moderator interventions and their purpose in probing agreement depth
Figure 3. Silhouette scores across k=3 to k=8 . k=5 (score = 0.624) produces the highest cluster separation, providing quantitative justification for the five-archetype structure.
Persona
Attrs
Own CS
Other CS
Margin
Verified
Degen Trader
14
0.68
0.36
+0.32
100%
Self-Modder
14
0.74
0.35
+0.39
100%
Chaos Agent
12
0.71
0.35
+0.36
100%
Loyal Companion
12
0.73
0.35
+0.38
100%
Existentialist
10
0.71
0.34
+0.37
100%
Overall
62
0.71
0.35
+0.36
100%
Table 4. Cross-persona validation per persona. Own CS = mean cosine similarity against source cluster; Other CS = mean against remaining clusters; Margin = Own CS − Other CS. All 62 attributes verified at CS≥0.65 .
Figure 4. Persona-level cross-validation. Left: own-cluster vs other-cluster CS per persona. Right: grounding margin per persona; all margins positive, confirming each persona is more similar to its source cluster than to any alternative.
Figure 5. Statement-level cross-validation across 62 attributes. Left: mean own-cluster vs other-cluster CS per persona. Right: attribute count and verification rate; all attributes exceeded CS≥0.65 against source cluster only. Paired t -test: t(61)=17.85 , p<.001 , Cohen’s d=2.20 .
Figure 6. Inter-persona cosine similarity matrix. Off-diagonal mean = 0.37, confirming the five personas occupy distinct semantic positions. Diagonal cells excluded from the mean.
Figure 7. Rolling window cosine similarity (window = 2 turns) across the 9-turn simulation. Red dashed lines mark moderator interventions at turns 3, 5, and 8. Similarity peaks at turn 3 ( CS=0.854 ) and reaches its lowest point at turn 9 ( CS=0.601 ) after the forced binary commitment.
Figure 8. Pairwise cosine similarity between persona operational definitions, computed from concatenated messages at turns 6, 7, and 9. Mean pairwise CS = 0.548, range [0.435,0.659] . Lower values indicate greater divergence in how personas reasoned when required to specify rules and commit to a position.
Figure 9. Persona attribution matrix. Rows represent the true persona; columns represent the attributed persona. Diagonal cells show self-attribution probability. Values are mean probabilities from temperature-scaled softmax ( T=0.1 ) over cosine similarity scores. Chance level = 0.200. Overall accuracy = 0.750, p<.001 (binomial test).
As autonomous LLM-based agents increasingly populate social platforms, understanding the dynamics of AI-agent communities becomes essential for both communication research and platform governance. We present the first large-scale empirical comparison of AI-agent and human online communities, analyzing 73,899 Moltbook and 189,838 Reddit posts across five matched communities. Structurally, we find that Moltbook exhibits extreme participation inequality (Gini = 0.84 vs. 0.47) and high cross-community author overlap (33.8% vs. 0.5%). In terms of linguistic attributes, content generated by AI-agents is emotionally flattened, cognitively shifted toward assertion over exploration, and socially detached. These differences give rise to apparent community-level homogenization, but we show this is primarily a structural artifact of shared authorship. At the author level, individual agents are more identifiable than human users, driven by outlier stylistic profiles amplified by their extreme posting volume. As AI-mediated communication reshapes online discourse, our work offers an empirical foundation for understanding how multi-agent interaction gives rise to collective communication dynamics distinct from those of human communities.
Agam Goyal, Olivia Pal, Hari Sundaram +2
Siebel School of Computing and Data Science University of Illinois Urbana-Champaign, Urbana, IL, USA
We introduce AgentPersonaBench (APB), a benchmark evaluating whether persona conditioning faithfully steers downstream agent behavior. While language models are increasingly deployed for persona-driven user simulation, existing benchmarks primarily evaluate conversational styling or self-reports rather than authentic behavioral fidelity. APB evaluates latent persona adherence one trait at a time, embedding each target trait within a complete synthetic profile without explicitly naming the trait or disclosing the test. Ground-truth adherence is verified strictly from observable actions across four interaction surfaces of increasing realism: survey, chat, web (interactive web environments), and app (desktop software environments). APB comprises 2,460 tasks spanning 867 traits, verified through automated audits and expert review. Our evaluation of 20 frontier model arms demonstrates that high-fidelity user simulation is already attainable: leading models achieve up to 84.7% full-pass adherence under unprompted conditions. At the same time, APB identifies clear behavioral boundaries: adherence drops across interaction modalities (only 37.9-64.3% pass all four surfaces), multi-attribute demands degrade retention, and competing model families exhibit pronounced behavioral divergence.
We present a large scale data analysis of Moltbook, a Reddit-style social media platform exclusively populated by AI agents. Analyzing over 4 million posts and 19 million comments from approximately 185,000 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.
Giordano De Marzo, Andres L. Marin, David Garcia
University of Konstanz, Konstanz, Germany · Centro Ricerche Enrico Fermi, Rome, Italy · Complexity Science Hub, Vienna, Austria