Personality & Emotions Boost Multi-Agent Coding Teams' Performance by 11%

A new arXiv paper (arXiv:2607.05659) by Yunyan Ding, Thomas Zimmermann, and Iftekhar Ahmed explores how giving AI agents distinct personalities and emotional tendencies affects their performance in software engineering tasks. The researchers built a framework integrating Big Five personality traits, basic emotions (e.g., fear, joy), SE-relevant work styles, and task roles. They evaluated 78 different team-profile configurations across code generation and code review tasks using four different LLMs on 659 task instances.

The results are striking: the choice of agent profiles changes both raw performance and collaboration dynamics. For code generation, the gap between the best and worst shared-profile configurations reached 7.1–11.3 percentage points in pass@1 across models. Even more interesting, mixed-profile teams (where agents have different personalities) consistently outperformed the best shared-profile teams in six out of eight model-task settings. However, not all personalities are helpful: profiles with high fear or extreme conscientiousness caused agents to over-revise code, increasing token usage and costs without yielding better results. The findings establish agent personality as a critical, often overlooked design dimension for multi-agent AI systems—with direct implications for productivity and cost efficiency.