cs.HCMar 3, 2026
SaveHow to Model AI Agents as Personas?: Applying the Persona Ecosystem Playground to 41,300 Posts on Moltbook for Behavioral Insights
Organizations: University of Vaasa, , Finland · Qatar Computing Research Institute, HBKU, , Qatar
Abstract
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 to . (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 .
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 against source cluster only. Paired -test: , , Cohen’s .
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 ( ) and reaches its lowest point at turn 9 ( ) 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 . 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 ( ) over cosine similarity scores. Chance level = 0.200. Overall accuracy = 0.750, (binomial test).