Agents with Feelings? Personality and Emotion in Multi-Agent Software Teams
This paper demonstrates that incorporating personality and emotion profiles into multi-agent LLM systems for software engineering significantly impacts team performance and collaboration efficiency, with mixed-profile configurations often outperforming shared ones while specific traits like fear and high conscientiousness can increase costs without consistent gains.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are building a software team, but instead of hiring real people, you are hiring a group of AI robots (called "LLM agents") to write code and check each other's work. Usually, when we set up these robot teams, we just tell them what to do: "You are the planner," "You are the coder," and "You are the reviewer."
This paper asks a different question: What if we also told them how to act?
Think of it like casting a movie. You know the script (the job description), but you also need to decide on the actors' personalities. Should the coder be a nervous perfectionist? Should the reviewer be a relaxed optimist? Or should everyone be the same?
The researchers ran a massive experiment to see if giving these AI robots specific "personalities" and "moods" actually changes how well they work together.
The Experiment: The "Personality Lab"
The researchers created a "personality recipe" for their robots using three ingredients:
- Big Five Traits: They picked three main personality types (like being very organized vs. messy, or very curious vs. practical).
- Moods: They assigned emotions like happiness, fear, anger, or sadness.
- Work Styles: They gave them specific habits, like "loves to try new things" or "prefers to stick to safe, known methods."
They tested these robots on two main jobs:
- Code Generation: Writing a program from scratch.
- Code Review: Reading code and finding mistakes.
They tested two team setups:
- The "Uniform" Team: Every robot in the group had the exact same personality and mood.
- The "Mixed" Team: Each robot had a different personality (e.g., a cautious planner, a happy coder, and a fearful reviewer).
The Big Discoveries
1. Personality Matters (A Lot)
Just like in a human office, the "vibe" of the team changed the results.
- The Gap: The difference between the best-performing personality mix and the worst-performing mix was huge. For code writing, the best teams solved 11% more problems than the worst teams. That's like one team finishing a whole extra week of work while the other team struggles.
- No One-Size-Fits-All: What worked for one AI model didn't work for another. A "fearful" personality might help a small, simple AI, but a "happy" personality might be better for a giant, complex AI. You can't just copy-paste a "perfect team" recipe; you have to tune it for your specific robot.
2. Mixed Teams Are Often Better
The researchers found that giving every robot the same personality (the "Uniform" team) was often a mistake.
- The "Specialized" Advantage: In most cases, the teams where each robot had a different personality performed better than the teams where everyone was the same. It's like a sports team: you don't want five goalkeepers; you want a mix of a striker, a defender, and a goalie. Different roles seemed to need different "personalities" to work best.
3. The "Fearful" Trap (More Work ≠ Better Results)
This was the most surprising finding. The researchers thought that if they made the robots more anxious or overly careful, they might do a better job.
- The Result: They did work harder, but not better.
- The "Fearful" Robot: When the team was assigned "fear" or "high conscientiousness" (being super careful), they revised their work constantly. They checked and re-checked the code over and over.
- The Cost: This made them use a lot more computer power (tokens) and take longer. But here's the catch: They didn't actually solve more problems. They just got stuck in a loop of "over-thinking" and re-doing things that were already correct. It's like a student who erases their answer 10 times because they are scared of making a mistake, but the answer was right the first time.
4. Mood Doesn't Make Them "Feel"
The paper clarifies that these robots don't actually feel fear or happiness. The "personality" is just a set of instructions (a prompt) that tells the AI how to behave.
- The Metaphor: Think of it like an actor putting on a costume. The actor isn't actually a cowboy, but they act like one because of the costume and the script. The AI isn't "scared," but it acts cautiously because the prompt told it to be "fearful."
The Bottom Line
If you are building a team of AI agents to do software work:
- Don't just assign roles; assign personalities. Who they are matters as much as what they do.
- Mix it up. Giving different robots different personalities often yields better results than making them all the same.
- Beware of "Anxiety." Making your AI too careful or fearful doesn't make it smarter; it just makes it slower and more expensive without improving the final product.
In short, designing an AI team isn't just about the software code; it's about the "character" of the team.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.