Simulating Teams with LLM Agents: Interactive 2D Environments for Studying Human-AI Dynamics

TL;DR

This work presents VirT-Lab, a web-based platform that uses LLM-based agents in interactive 2D environments to simulate human team dynamics. It advances beyond traditional ABM and RL by integrating scalable agent-based modeling with spatial reasoning, memory, and dynamic communication, while offering a no-code, visualization-rich workflow for researchers and practitioners. The paper contributes four design goals, a detailed system architecture, and a comprehensive evaluation including ground-truth, scalability, ablation, and user studies demonstrating both realism and usability. Overall, VirT-Lab provides a versatile testbed for exploring how environments shape coordination and collaboration in human–AI teams, with open-source potential and broader implications for designing responsible AI-assisted team simulations.

Abstract

Enabling users to create their own simulations offers a powerful way to study team dynamics and performance. We introduce VirTLab, a system that allows researchers and practitioners to design interactive, customizable simulations of team dynamics with LLM-based agents situated in 2D spatial environments. Unlike prior frameworks that restrict scenarios to predefined or static tasks, our approach enables users to build scenarios, assign roles, and observe how agents coordinate, move, and adapt over time. By bridging team cognition behaviors with scalable agent-based modeling, our system provides a testbed for investigating how environments influence coordination, collaboration, and emergent team behaviors. We demonstrate its utility by aligning simulated outcomes with empirical evaluations and a user study, underscoring the importance of customizable environments for advancing research on multi-agent simulations. This work contributes to making simulations accessible to both technical and non-technical users, supporting the design, execution, and analysis of complex multi-agent experiments.

Figures (11)

• Figure 1: Scenario description interface in the VirT-Lab system. (A) Users enter the description of the simulation scenario, outlining the mission and its objectives. (B) The Enhance Description button allows users to refine and enhance their input with AI-generated improvements. (C) The system detects clear scenario goals based on the user's scenario description. (D) Users can review these validated objectives before proceeding.
• Figure 2: Overview of the VirT-Lab interface for simulation setup.
• Figure 3: The agent configuration interface of VirT-Lab. The system allows users to customize agents by defining demographic attributes, backgrounds, and personality traits. (A) Agent cards present each agent with the agent's name and role in the simulation. (B) Agent background and experience are presented as narrative elements that shape the agent’s behavior. (C) Displays the number of agents generated and currently active in the simulation. (D) The "Add Agent" button allows users to expand the team by creating new agents. (E) View allows users to access agent details, including agent demographics and personality. (F) Agent personality trait summarizes the agent’s Big Five personality profile, (G) The personality background and experience. (H) The "Edit Profile" button provides options to modify demographic information, background, and personality traits, and (I) The personality chart visualizes behavioral tendencies.
• Figure 4: Overview of the simulation environment created in VirT-Lab. (A) The map displays the environment's physical constraints (e.g., walls), interactive entities (e.g., victims), and non-interactive entities (e.g., trees). (B) Two regions of the map environment (R1 and R2). Agents navigate the environment based on spatial constraints. (C)VirT-Lab uses a partitioning algorithm to recursively subdivide the regions into a graph $\mathcal{G}$. The leaf nodes of the graph represent navigable sub-regions. (D) The map is encoded as a binary matrix $\mathcal{M}$ that agents use to perceive their surroundings. Each environment's matrix $\mathcal{M}$ is dynamically updated throughout the simulation to reflect environmental changes and agent actions. Number '0' represents limits for the agents' movements (e.g., walls) while each agent's ID is located in the matrix. In this case, the agent with ID 4 is added to the matrix representation.
• Figure 5: Simulation execution interface of VirT-Lab. The system integrates different panels to display agent behavior and team dynamics. (A) The scenario overview provides the mission context initially defined by the user. (B) Narrative panel presents descriptive text updates for each active agent in the simulation, generated by the system to reflect the current state of the environment. (C) Timeline tracks events across simulation steps; at each timestep, users can review scenario updates, environmental changes, and agent actions. (D) Event details show agents' activity, commands, and outcomes. (E) A 2D grid map visualizes the environment layout and agent events as they unfold during each timestep. (F) Simulation controls allow users to pause and revisit actions. (G) The data logging button enables users to download structured logs of the simulation in JSON format. (H) Simulation analysis provides the users with qualitative insights into the simulation. (I) The communication panel records conversations among agents. These components provide users with a comprehensive view of the simulation state, actions, and interactions.