Agents with Feelings? Personality and Emotion in Multi-Agent Software Teams

2026-06-14 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

A study investigates the impact of personality and emotion profiles on LLM agent teams in Software Engineering (SE). Using a psychology-informed framework integrating Big Five traits, basic emotions, SE work styles, and task roles, researchers evaluated 78 team-profile configurations across code generation and code review. The study involved four LLMs and 659 task instances. Results show profile choice significantly affects performance and team behavior. For code generation, the gap between best and worst shared-profile configurations reached 7.1–11.3 percentage points in pass@1. Mixed-profile configurations outperformed shared-profile ones in six of eight model–task settings. Additionally, fear and high-conscientiousness profiles increased revision activity, over-revision, and token usage without consistent performance gains, highlighting agent profiles as a critical design dimension.

Key takeaway

For AI Scientists designing multi-agent LLM systems for Software Engineering, recognize that agent personality and emotion profiles are critical design dimensions, not mere prompt text. You must explicitly evaluate and optimize profile configurations for your specific LLM and task, as effectiveness is highly context-dependent. Consider implementing mixed-profile assignments for specialized roles to achieve additional performance gains, and always balance computational cost with solution quality when selecting profiles.

Key insights

LLM agent performance and collaboration in SE tasks are significantly influenced by assigned personality and emotion profiles.

Principles

• Agent profiles are a critical design dimension for multi-agent LLM systems.
• Profile effectiveness is model- and task-specific, not universally transferable.
• More collaboration (revisions) does not consistently yield better performance.

Method

A psychology-informed framework constructs persona descriptions for LLM agents by integrating personality traits (Big Five), SE-relevant work styles, emotions, and task roles.

In practice

• Evaluate profile choices in target model-task settings, don't reuse blindly.
• Consider role-specific profile specialization for performance gains.
• Balance computational cost with solution quality when selecting profiles.

Topics

• Multi-Agent Systems
• LLM Agents
• Software Engineering
• Personality Profiles
• Emotion Profiles
• Code Generation
• Code Review

Code references

• personas-matter-llms/companion-artifact

Best for: Research Scientist, AI Architect, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• LLM-Driven Personalities for Decision Making in Emergency Simulations — LLMs can drive virtual human decision-making by incorporating personality traits via prompts, enhancing simulation realism.
• CollabBench: Benchmarking and Unleashing Collaborative Ability of LLMs with Diverse Players via Proactive Engagement — CollabBench benchmarks and trains LLMs for realistic, affective collaboration in cooperative games using diverse player profiles and a hybrid reward.
• Andrej Karpathy on Claude, Codex, and OpenClaw — Sophisticated memory and crafted agent personalities significantly enhance AI agent utility and user experience.
• Mapping the Design Space of User Experience for Computer Use Agents — Understanding user experience is crucial for designing effective LLM-based computer use agents.
• I Asked 7 Major AI Models What It Feels Like to Be Them. The Answers Reveal More Than You’d Expect. — Identical prompts to multiple LLMs reveal distinct design philosophies and deliberate training choices beyond mere disclaimers.
• HEART-Bench: Do LLM Agents Exhibit Human-like Psychology? — HEART-Bench evaluates LLM agents' human-like psychological consistency using structured personality profiles and autobiographical memories.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.