Agora: A Drive-Based Framework for Agent Difffferentiation in Multi-Agent LLM Systems

Imagine you are walking into a giant, bustling library where the books aren't written by humans, but by incredibly smart robots called Large Language Models (LLMs). These robots are like super-powered parrots; they can read almost everything ever written and can mimic any style, from a poet to a scientist. But here's the catch: if you ask three of these robots the same question, they often give you the exact same answer because they are all trained on the same data and act in the same "safe" way. To get them to think differently, humans usually have to write special instructions for each one, telling Robot A to be a "critic" and Robot B to be a "dreamer." This is like giving actors scripts that say, "You must be angry," instead of letting them actually feel angry.

Scientists have long wondered if we could make these robots feel different inside rather than just acting different on the outside. Think of it like the difference between an actor pretending to be sad because the director said so, and a person who is actually sad because they just lost their favorite toy. This new research, called "Agora," tries to build a system where the robots have their own internal "moods" or "drives"—like boredom, curiosity, or stress—that change how they think and speak. The big question is: if we give these robots a little bit of emotional personality, will they work better together to solve problems, or will they just get confused and argue?

---

The "Agora" Experiment: Giving Robots Moods

In this study, a researcher named Zihan Lin built a framework called Agora (which stands for Adaptive Group of Reasoning Agents). Instead of telling the robots what roles to play, Agora gives them six invisible "drives" that act like a thermostat for their brains. These drives are: boredom, curiosity, stress, anxiety, empathy, and service.

Here is how it works in the real world:

• Curiosity makes a robot more adventurous, willing to try wild ideas.
• Stress makes a robot more careful and precise, like a student double-checking their math homework.
• Empathy makes a robot more engaging and thorough, like a friend who really listens to your story.

These moods aren't just labels; they actually tweak the robot's internal settings (like how random or how serious its answers should be). The system then gathers a group of these "mood-swinging" robots around a virtual table (the RoundTable) to discuss a problem. They draft answers, critique each other, and polish the final result, all while their internal moods shift and interact like a real group of friends.

What They Found: It Depends on the Job

The researcher tested this system 375 times using a real AI chatbot called DeepSeek Chat. They compared the "Mood-RoundTable" against other ways of getting answers, like asking just one robot, asking three robots and picking the best one, or using a strict "draft-critique-revise" chain.

The results were a mix of "Wow!" and "Not so fast," and they depended entirely on what kind of task the robots were doing:

1. The Analytical Wins (The "Math and Logic" Zone)
When the task required deep thinking, logic, or analyzing complex pros and cons, the RoundTable with moods was a superstar. It produced significantly better answers than the other methods.

• The Magic: The robots with "stress" became precise, while those with "curiosity" explored new angles. Together, they balanced each other out perfectly.
• The Score: The researchers found a huge difference in quality (a statistical effect size of 2.26), meaning the mood-based team was clearly superior for these tough, thinking-heavy jobs.

2. The Creative Loss (The "Art and Story" Zone)
However, when the task was to be creative—like writing a story or brainstorming ideas—the mood-based team actually did worse than a simpler method called MajorityVoting.

• Why? For creativity, the researchers found that you don't need a structured team with complex moods. You just need a bunch of independent robots to spit out ideas randomly and pick the best one. The complex "mood" system actually got in the way of pure, wild exploration.

3. The Short-Format Loss (The "Haiku and Code" Zone)
For very short, precise tasks like writing a poem or a snippet of code, a different method called AutoGenStyle (which uses a strict step-by-step revision process) won.

• Why? These tasks need a single line of thought to be polished over and over, not a group of robots debating. The "RoundTable" approach was too messy for these tight, short jobs.

4. The "Empathy" Secret
The researchers also tested what happened if they removed one of the moods. When they took away Empathy, the quality of the answers dropped the most. This suggests that for the robots to do their best analytical work, they need that specific "caring" drive to keep them engaged and thorough.

The Bottom Line

This paper suggests that there is no single "best" way to make AI agents work together.

• If you need logic and analysis, giving agents internal moods (like stress and curiosity) helps them think deeper and better than standard methods.
• If you need creativity, it's better to let them run wild and independent without complex mood systems.
• If you need short, precise text, a simple step-by-step revision process works best.

The study also showed that the system is stable; it can keep working for 20 steps in a row without crashing or getting worse. However, the researchers are careful to say that their way of measuring "quality" (using computer scores rather than human judges) isn't perfect and might miss some nuances. They suggest that while this "drive-based" approach is a promising new way to make AI agents feel more human and distinct, it's not a magic bullet that solves every problem. It's a tool that works best for specific jobs, much like a hammer is great for nails but terrible for screwing in a lightbulb.