Personality prompting in multi-agent LLM teams shows task-dependent effects on performance

A new arXiv preprint examines how personality prompting in large language models (LLMs) influences multi-agent team performance across three task domains: structured coding, open-ended research collaboration, and competitive bargaining.

The authors manipulated agreeableness traits in frontier LLMs and observed that communication styles shifted predictably—low agreeableness produced adversarial language, while high agreeableness yielded cooperative behavior—but these shifts did not uniformly translate to task outcomes.

In structured coding tasks, large communication shifts induced by low agreeableness had little effect on milestone completion, suggesting task structure can buffer against behavioral manipulation.

By contrast, the same agreeableness manipulation substantially degraded performance in open-ended research collaboration and competitive bargaining, indicating that unstructured or competitive environments are more sensitive to personality-induced communication dynamics.

The findings highlight the limits of personality manipulation as a tool for controlling multi-agent LLM teams and underscore the importance of aligning agent design with the structural demands of the task.