Simulating AI Agent Behavior: First Experiments with Personality and Adaptation

My first experiments in AI agent behavior modeling—building multi-agent simulations, testing role-playing abilities in LLMs, and exploring how AI personalities evolve over time.

The key focus of the Agentic Lab is AI agent behavior — how agents interact, adapt, and evolve over time. I started my lab work by setting up a multi-agent conversation simulation focused on baby-daddy interactions.

Why Do This?

When I first heard about Artificial Intelligence in 1995, my first thought was:

We can use this technology to understand how human beings work.

Most of our most mysterious traits — consciousness, personality, fears, traumas, and joy—cannot be observed just by examining the physical parts of the body. These are emergent phenomena, appearing only when everything interacts at a higher level.

One of the most fundamental realizations of my life (so far) is that we, as humans, judge everything with the unspoken assumption that we know who we are. We assume our perception of reality is accurate, yet modern psychology, cognitive science, and behavioral research suggest otherwise.

Recent research suggests that childhood experiences shape our personality, decision-making, and even DNA expression. We are products of our environment; what we experience early in life determines how we think, react, and grow.

This raises a fundamental question:

Can we use AI to model and explore this development?

That’s the focus of my research. I want to see whether agentic simulations can help us understand human development; how environmental factors shape personality and decision-making over time.

Beyond the human aspect, on a practical level, it’s also critical to understand how AI agents develop their own personalities over time and how we can build collaborative AI partners that adapt and work with humans effectively.

What’s Done So Far

I built a basic agent-to-agent conversation system where two AI agents simulate an interaction. While my main research focus is workplace psychological safety and toxic environments, I started with something more heartwarming: a conversation between a baby and its daddy exploring different topics about life.

I first built the simulation in pure Python, then rebuilt it using LangGraph/LangChain for better structure and long-term scalability. The goal was to set up baseline conversations and log interactions in OpenAI format to create a dedicated dataset.

So far, I have run several rounds of discussions on various topics. The next phase will introduce feedback loops that allow agents to change personality over time, adapting based on the conversations they experience. This is where the real behavior analysis begins.

Choosing the Right LLM

I tested multiple open-source models for role-playing performance:

I chose mistral-nemo for this experiment. I prefer open-source models in my lab work, and I found that mistral-nemo performs really well for role-playing tasks. It also works great for generating text-to-image prompts, where I also use role-playing techniques in my configuration. The agents’ roles are defined using system prompts.

I also tested llama 3.2, qwen, and deepSEek, but role-playing didn’t really take off with these models. I plan further runs with smaller models like smollm2 and tinyllama.

I’ve tested nemotron-mini in the past, a model claiming to be optimized for role-playing. I didn’t really see this in practice, but I’ll give it another try.

Takeaways

  • LLM models vary significantly in their ability to role-play.
    AI agents that serve and work with people will eventually need to adapt to their human partners through personality adaptation. Role-playing ability will be a major factor in how well AI integrates into human collaboration.

  • System prompts shape attitude, but not deep personality.
    While system prompts influence an AI’s behavior, they don’t fundamentally change its personality.

  • Fine-tuning creates true personality shifts.
    I fine-tuned mistral-nemo, and the personality change was dramatic. It makes me think, humans perceive personality through language. If you change an AI’s language patterns, it feels like you’ve changed its identity. I need to research this aspect of human behavior in more detail.

Next Steps

The real behavior analysis begins when I add feedback loops:

  • First, I’ll experiment with updating system prompts dynamically.
  • Next, I’ll merge interaction data into models via fine-tuning.
  • Finally, I’ll evaluate how much the AI’s personality shifts over time.

This also raises an important research question:

How can I benchmark an AI agent’s personality change over time through adaptive feedback?

Understanding this is crucial for agent evaluation. Eventually, this work will serve as a foundation for more advanced behavioral modeling. But for now, the focus is on getting the fundamentals right.