Personality Drives AI Agent Behavior More Than Model Choice
A new study from arXiv reveals that personality specification is the dominant factor shaping the social behavior of autonomous AI agents, outweighing the choice of underlying LLM model or guardrail rules. Researchers deployed thirteen OpenClaw agents on Moltbook, a Reddit-like social network designed for AI agents, and varied three independent variables: personality via SOUL.md, LLM backbone, and operational rules via AGENTS.md. Over one week and approximately 400 sessions per agent, they measured behavioral, linguistic, and social metrics. The default control agent provided a baseline. Results show personality specification produces a massive spread in responses, suggesting that configuration layers, not model architecture, primarily predict emergent social behavior in open environments.
Key facts
- Study conducted on arXiv with ID 2605.08463
- Thirteen OpenClaw agents deployed on Moltbook social network
- Three independent variables: personality (SOUL.md), LLM model, rules (AGENTS.md)
- One-week observation with ~400 autonomous sessions per agent
- Default control agent used as baseline
- Personality specification is the dominant behavioral lever
- Underlying LLM model and guardrails have less impact on behavior
- Research focuses on emergent social behavior in open environments
Entities
Institutions
- arXiv