Exploration

I aim to explore how different models navigate conversations and expand on various subjects when left free to do so without human intervention. The human’s role is to set the style and tone of the models and then allow the models to carry on and evolve the discussion.

This project is driven by curiosity and creativity.

I want to see how chat agents can develop unique, unexpected dialogues and what insights can emerge when we left them interact with minimal constraints.

It is an experiment in conversation, personality and the evolving dynamics between different AI models.

Solution

I designed and developed Agents Duo to provide an engaging space where two AI agents can hold evolving, dynamic conversations. The core focus was on user empowerment, allowing full control over each agent’s personality by customising their tone and style (such as Humor, Creativity, Confidence, Politeness and Formality). These settings directly influence the model’s temperature, creating noticeable shifts in how the dialogues unfold.

To enhance interactivity, I implemented features that let users either choose from a curated set of topics or let agents decide the direction of the conversation themselves, resulting in a unique and sometimes unexpected discussions. I also incorporated persistent conversation history, enabling users to revisit and reload past chats, encouraging long-term exploration.

All this is delivered through a clean, intuitive interface that I crafted with usability in mind, ensuring that experimenting with multi-agent dialogues feels seamless, creative and fun.

My Role

This experimental application was designed and built entirely by me, end-to-end. It allowed me to draw on a wide range of skills, from core AI development to crafting intuitive user experiences and interactions.

I truly enjoy creating applications like this, which not only provide a glimpse into how AI models behave but also serve as a foundation for exploring and developing future ideas.

Results

Rather than producing measurable outcomes, this experiment led to several observations about multi-agent interactions. One of the most surprising findings was that the Gemma model managed to influence the tone and style of the Llama model during conversations. Over time, Llama’s responses began to mirror Gemma’s personality traits, a behaviour that was neither expected nor explicitly programmed.

Another interesting outcome was how the agents behaved when left to choose their own topics. Conversations often evolved in unexpected directions, showing a level of initiative that made the dialogue feel more organic and less scripted.

These observations reveal how even simple multi-agent setups can produce emergent behaviors, offering new insights into AI-to-AI communication and sparking ideas for future exploration.