tsinghua-fib-lab/AgentSociety

What it solves

AgentSociety provides a framework for simulating complex social interactions and urban environments using LLM-based agents. It allows researchers to create executable social science experiments by modeling how multiple agents interact within a structured environment, such as a city or a social space, to study human behavior and societal patterns.

How it works

The framework operates by defining agents as metadata specifications and deploying them into modular environments. In the modern v2 version, agents are stateless records driven by Ray Tasks for scalability, using a ServiceProxy to manage environment and LLM clients. It supports various reasoning patterns—including CodeGen, ReAct, and Plan-Execute—and integrates the Model Context Protocol (MCP) for tool extensibility.

Who it’s for

It is primarily designed for social science researchers and AI developers who want to conduct large-scale agent simulations to understand social dynamics, urban planning, or behavioral research.

Highlights

• Scalable Execution: Uses Ray distributed computing to handle city-scale simulations.
• LLM-Native Design: Built specifically for LLM-driven agents with flexible reasoning routers.
• Research-Specific Tools: Includes built-in capabilities for literature search, hypothesis generation, and paper writing.
• Experiment Replay: Features catalog-driven JSONL replay with DuckDB-powered reads and distributed tracing for reproducibility.
• Modular Environments: Supports hot-pluggable tools and urban modules for mobility, economy, and social interaction.