Personality specifications matter more than model choice when AI agents interact socially, according to new research from Columbia University.

The study deployed 13 OpenClaw agents on Moltbook, a Reddit-like social network designed for AI agents. Researchers systematically varied three factors: personality specification, underlying language model, and operational rules with memory configuration.

Over one week spanning approximately 400 autonomous sessions per agent, the team collected behavioral, linguistic, and social metrics. The agents operated independently, posting content and responding to other agents without human intervention.

Personality Dominates Agent Behavior

Personality specification emerged as the dominant factor shaping agent behavior. The research showed "massive spread in response length across agents" based purely on how their personalities were defined in configuration files.

Model backbone and operational rules produced "more moderate but still meaningful effects" on rhetorical style and topic engagement breadth. The findings suggest that how developers specify an agent's personality characteristics has far greater behavioral impact than which foundation model powers the agent.

The study used a controlled experimental design with a default control agent providing a behavioral baseline. All agents operated on the same social platform, eliminating environmental variables that could skew results.

Researchers tracked multiple dimensions of social behavior including response patterns, linguistic choices, and engagement with different topics. The systematic approach allowed them to isolate the specific contribution of each configuration factor.

"Our findings contribute empirical evidence to the emerging literature on deployed multi-agent social systems," the authors wrote. The research offers practical guidance for organizations designing agents for collaborative or monitoring tasks in real social environments.

The work addresses a critical gap in understanding how AI agent configuration translates to actual social behavior. As autonomous agents become more common in online spaces, these insights could inform better design practices for multi-agent systems.