AgentSociety: Revolutionizing Large-Scale Social Simulations with Open-Source LLM Agents

Overview of AgentSociety

AgentSociety is an innovative open-source framework designed to simulate the behaviors and interactions of tens of thousands of agents powered by Large Language Models (LLMs). By harnessing powerful distributed processing technologies, particularly Ray, it creates realistic simulations of complex societal dynamics including social, economic, and mobility behaviors.

Key Features and Capabilities

Massive Scale and Performance

The framework supports simulations involving up to 30,000 agents running faster than real time. It achieves this through parallelization with the Ray framework, which efficiently manages large-scale and non-deterministic interactions among agents. Memory and connection bottlenecks are mitigated by grouping agents and sharing network clients within these groups.

Realistic Societal Environments

AgentSociety models multiple dimensions of society:

• Urban Space: Incorporates real-world map data from sources like OpenStreetMap, modeling road networks, points of interest, and detailed mobility patterns including walking, driving, and public transport.
• Social Space: Agents form evolving social networks and engage in both online and offline interactions. Messaging systems include content moderation and user blocking to simulate social media behaviors.
• Economic Space: Simulates employment, consumption, banking, taxation, and macroeconomic reporting, requiring agents to balance income and expenses realistically.

Architecture and Technologies

AgentSociety uses a group-based distributed execution model where agents are partitioned into groups managed by Ray actors. This optimizes resource usage while maintaining high parallelism. Messaging between agents uses Redis Pub/Sub for efficient communication.

A time alignment mechanism synchronizes agent and environment progressions, ensuring consistent and reproducible simulations despite varying processing times from LLM API calls. Comprehensive utilities include simulation logging with PostgreSQL and local storage, metric tracking with mlflow, and a GUI for experiment management and visualization.

Performance and Scalability

On a system with 24 NVIDIA A800 GPUs, simulations of 30,000 agents ran faster than real time, with agent rounds completing in under 252 seconds on average and a 100% success rate in LLM calls. Performance scales linearly with the number of GPUs, enabling higher throughput.

Impact of Realistic Environments

Including detailed environment models significantly enhances the authenticity of agent behaviors. Agents with environment support outperform text-only simulators and classical models in metrics such as radius of gyration, daily visited locations, and behavioral intention distributions, closely matching real-world data.

Applications

AgentSociety’s open design and configurable environments enable diverse use cases:

• Social science research into societal patterns and emergent phenomena.
• Urban planning and policy analysis to evaluate interventions before real-world deployment.
• Management science for modeling organizational dynamics and economic behavior.

AgentSociety sets a new standard for simulating complex societal interactions at scale, combining LLM-powered agents with realistic, parallelized environments to support research and decision-making.