Irrational Generative Agents – Capstone Project

Project Overview

This capstone project presents Irrational Generative Agents, a cognitive agent framework designed to simulate human-like reasoning and decision-making, including both rational and irrational behaviors.

The system aims to bridge gaps between psychology, artificial intelligence, and agent-based simulation by modeling memory, emotion, motivation, and cognitive bias within autonomous agents operating in a simulated world.

Background and Motivation

Modeling realistic human behavior remains a core challenge in both psychology and AI.

Key challenges motivating this project include:

Human-subject research is costly, slow, and ethically constrained
Existing virtual agents lack deep psychological grounding
Most agents fail to model persistent irrational biases or memory-driven behavior

This project addresses these limitations by creating agents capable of bounded rationality, emotional influence, and long-term behavioral evolution.

Framework Overview

The framework integrates dual-process cognitive theory:

System 1: Fast, intuitive, heuristic-driven reasoning
System 2: Slow, analytical, deliberative reasoning

Core Characteristics

Python-based cognitive and decision-making model
Unity-powered 2D simulation environment
Customizable scenarios with real-time observation
Modular architecture supporting experimentation and extension

Key Contributions

Design of an irrational agent architecture integrating:
- Sensory input
- Motivation
- Emotion
- Cognitive biases
A modular framework for studying bias-driven decision-making
A lifelike agent loop enabling long-term behavioral growth
A visually grounded simulation world for real-time analysis

Project Management and Structure

Agile Scrum methodology
Project duration: August 2024 – May 2025
Three development sprints:
1. Exploration & Design
2. Coding & Development
3. Validation & Testing

Each sprint focused on clearly defined deliverables to ensure steady and measurable progress.

Requirements and Features

Core Requirements

Support for both System 1 and System 2 reasoning
Simulation of psychological biases
Persistent memory and personality modeling

Implemented Features

2D town simulation environment
Agent personality and emotion modules
Social interaction dynamics

Excluded (v1.0)

Dynamic map generation
Real-time direct user control of agents

Use Cases and Team Collaboration

Primary Use Cases

Simulation initialization
Autonomous agent behavior cycles
Observation data export

Team Workflow

Weekly team meetings
Biweekly supervisor check-ins
Tooling:
- GitHub (version control)
- Slack (communication)
- OneDrive (documentation)

Methodology and Literature Review

The framework draws inspiration from:

Generative Agents
Humanoid Agents
Evolving Agent systems

Addressed limitations include:

Lack of persistent irrational biases
Weak grounding of internal psychological states

Prompting Strategy

Combination of zero-shot and few-shot prompting
Structured prompt templates to stabilize LLM outputs

Memory and Decision-Making Logic

Memory System

Short-term memory: recent events and interactions
Long-term memory: consolidated knowledge
Memories are managed using an impact index

Decision-Making Flow

System 1 handles quick, low-cost decisions
System 2 manages planning, evaluation, and conflict resolution
Memory influences reasoning depth and behavioral consistency

Evaluation Strategy

Evaluation compares agent decisions against human participants.

Metrics Used

Kullback–Leibler Divergence for decision distribution similarity
Parallel decision-making tasks for agents and humans
Scenario-based behavioral comparison

Study Metrics for Agent Behavior

Human-choice prediction accuracy
Five personality dimensions scored from 0–10
Behavioral maturity:
- Low
- Medium
- High
Evaluation of n = 7 diverse personality profiles
Bias-awareness isolation through controlled prompting

System Architecture Overview

Business Architecture

Task configuration → cognitive processing → evaluation

Application Architecture

Perception module
Planning module
Bias evaluation module
Memory module
Action execution module

Agents are autonomous entities with personality, emotion, memory, and goals.

Agent Internal Architecture

Key components:

Stimulus processing
Dual-system decision-making
Action execution pipeline

Modules

Plan Module: generates candidate plans
Evaluation Module: scores plans using bias and memory relevance
Memory Module: ensures behavioral consistency over time

Cognitive Processing Pipeline

Environmental perception and interpretation
Semantic compression for memory and reasoning
Dynamic switching between System 1 and System 2
Context-aware behavioral variation

Inter-Agent Communication

Agents communicate indirectly via the environment:

Map tiles store interaction traces
Agents observe tiles within perceptual radius
Enables emergent social behavior

Interaction types include:

Entity perception
NPC interactions
Environmental reminders

Daily Behavior Cycle

Agent initialization (personality & emotion)
Environmental perception
Intent generation
Action planning and execution
End-of-day memory consolidation and trait adjustment

Action Planning Cycle

Inputs:
- Environmental perception
- Daily goals
- Internal emotional state
Plan categories:
- THINK
- CHAT
- MOVE
- INTERACT
Probabilistic plan selection
Post-action internal state updates

Frontend Architecture

React: UI, menus, HUD
Phaser: 2D simulation and interaction
Socket.IO: real-time frontend–backend communication
Supports multiplayer state synchronization

Technology Stack

Frontend

React
Phaser

Backend

Python
LangSmith (LLM integration)

Communication

Socket.IO

The system is modular and designed for scalability.

Evaluation Metrics

KL Divergence
Top-k behavioral overlap
Big Five personality shift
Behavioral maturity indicators:
- Multi-step planning
- Emotional regulation consistency

Experimental Setup

Web-based sandbox environment
26 agents with distinct personalities
Full logging of actions, emotions, and decisions
Bias-focused study using a subset of 3 agents
Human participants provide baseline comparisons

Results and Analysis

Key findings:

Mixed System 1 + System 2 reasoning aligns best with human behavior
Personality traits remain mostly stable
Behavioral maturity varies across agents
Emotional regulation improves over time, with occasional bias spikes

Modular Bias Evaluation

Clear behavioral differences between biased and unbiased agents
Bias-aware agents:
- Show richer narratives
- Engage in deeper conversations
- Exhibit stronger emotional tone
Bias significantly impacts social dynamics and decision-making

Limitations and Future Work

Current Limitations

No self-calibration of goal–behavior alignment
Static heuristics and personality filters
Fixed perceptual and interaction radii

Future Directions

Adaptive self-evaluation mechanisms
Dynamic personality evolution
More realistic social modeling
Systematic large-scale evaluation

Conclusion

The Irrational Generative Agents framework successfully simulates bounded and irrational human decision-making.

Agents exhibit diverse, personality-driven behavior
Memory integration enables long-term evolution
Emotional volatility decreases over time
The architecture provides a strong foundation for future research into cognition and behavior