[BETA] Social Agents

NO LLMs  /  NO CHATGPT  /  NO SCRIPTED CONVERSATIONS

Socially-Driven Digital Agents is an interactive research simulation exploring how personality, experience, and reinforcement learning can produce emergent social behavior in autonomous agents.

Rather than using language models, each agent operates through a hybrid Finite State Machine + Reinforcement Learning architecture. Agents perceive their surroundings, make decisions, interact with others, and learn from the outcomes of those interactions.

01   THE SIMULATION

Place a population of autonomous agents into a shared environment and watch their social behavior unfold in real time.

Each agent has its own personality traits, relationships, experiences, and preferences. These differences influence how agents approach others, respond to social situations, and decide what to do next.

• Personality-driven behavior
• Reinforcement learning from experience
• Individual relationships and affinity
• Social perception and proximity
• Self-organizing conversation groups
• Emergent cooperation, avoidance, and competition
• Persistent social patterns and group formation

There is no predefined story controlling who interacts with whom. Social structures emerge from the decisions of individual agents.

02   HOW AGENTS THINK

PERSONALITY

Agents are assigned personality-like traits based on the Big Five personality framework. Traits influence sociability, exploration, social preferences, interpersonal distance, and responses to different situations.

FINITE STATE MACHINE

The FSM provides an interpretable behavioral backbone. Agents can idle, seek social interactions, approach other agents, participate in conversations, work, seek food/shade, and disengage.

REINFORCEMENT LEARNING

Agents learn from experience rather than following completely fixed behavior. Interaction outcomes influence future decisions, allowing agents to gradually develop different social strategies.

03   SOCIAL EMERGENCE

Individual decisions can produce behaviors that were never explicitly programmed as outcomes.

Agents may repeatedly seek out people they have had positive interactions with. Others may avoid low-reward encounters. Groups can form around compatible personalities, relationships can strengthen through repeated interaction, and social clusters can change as agents learn.

Individual behavior → interaction → experience → adaptation → emergent social structure

04   WHY NO LLMs?

This project is not designed to simulate conversation through natural language generation.

The research focuses on social decision-making and emergent behavior. An agent does not need to generate a sentence to demonstrate social behavior. It can choose who to approach, how long to interact, whether to remain in a group, and whether a previous interaction should influence its next decision.

This makes the underlying behavior observable, interpretable, and measurable.

05   RUN THE EXPERIMENT

Experiment with different populations and observe how changing the conditions changes the resulting social dynamics.

• Compare different personality distributions
• Change agent density
• Observe conversation groups form and dissolve
• Track relationships and social clusters
• Watch learned preferences develop over time

The simulation records encounters, affinities, group formation, learning behavior, and other social metrics for analysis.

06   THE RESEARCH

Socially-Driven Digital Agents: A Hybrid Personality-Driven Architecture for Realistic Multi-Agent Social Interaction

This project investigates whether combining personality-driven finite state machines with reinforcement learning can produce more socially coherent multi-agent behavior without relying on large language models or fully scripted social interactions.

Every action done is learned. 

Interested in reading the paper? Stay tuned as it is getting archived!

Research project by Ian Wu Stetson University 2026