AgentSociety

AgentSociety simulates a society of people, each played by a large language model (LLM), inside an environment with rules of its own.^1,4 It comes from the FIB Lab at Tsinghua University in Beijing, and both papers are led by Jinghua Piao, with Yong Li as last author.^2,3,4

The first version, described in February 2025, is a city simulator.^1,3 It has modules for getting around a city, for an economy and for social ties, and runs at scale on Ray, a framework for spreading computation across many machines.¹ The paper generated social lives for over 10,000 agents and about 5 million interactions.³ It used them for five experiments: polarization, the spread of inflammatory messages, a universal basic income, the shock of a hurricane, and urban sustainability.³ It ran surveys, interviews and interventions inside the simulation.³ The authors report that the outcomes lined up with real-world experimental results.³

Version 2, described in mid-2026 and now the recommended package, widens the scope.^1,4 The same software runs two kinds of agent: simulated participants who respond inside a configurable social setting, and AI "social scientists" that search the literature, propose hypotheses, design the experiment, run it, interpret it and draft the paper, with a human researcher directing.⁴ The paper presents seven illustrative studies, from laboratory-style experiments through social media dynamics to urban scenarios, and says they reproduced the main qualitative patterns of earlier studies and showed informative deviations.⁴ Agents can reason in several styles, and runs can be replayed.¹

Both versions install with pip and need Python 3.11 or later and an API key for OpenAI, Anthropic or any provider supported by the litellm library.¹ Version 1 stays available as a legacy package.¹ A web front end and a VS Code extension are in the repository.¹ Version 2 had release 2.9.3 in September 2026, and commits continued into October.^5,6 The code is Apache-2.0 except for a folder marked commercial in the version 1 package.¹ The agreement with past findings is the authors' own report; the abstracts describe no independent evaluation.^3,4