AgentSociety 2 automates end-to-end social science with dual AI agents

AgentSociety 2 tackles a fundamental gap in automated social science: existing systems either assist isolated tasks or treat AI agents as passive subjects, leaving the research workflow disconnected from the simulated society. The new environment solves this by coupling two types of LLM agents in a single runtime: AI social scientists that coordinate the entire research process (hypothesis generation, experiment design, simulation execution, result interpretation, and manuscript writing) and silicon participants that generate behavioral responses within configurable social environments. This dual-role design turns abstract hypotheses into auditable agent behaviors, environment rules, and measurable outcomes, enabling a truly end-to-end workflow.

In seven illustrative studies, the team demonstrated AgentSociety 2's versatility across scales—from micro-level laboratory experiments (replicating classic social psychology findings) to meso-level social media dynamics and macro-level urban simulations. The system not only reproduced major qualitative patterns from prior studies but also identified informative deviations, suggesting new avenues for research. By preserving human researchers' high-level agency while delegating procedural orchestration to AI, AgentSociety 2 provides a controllable, human-in-the-loop infrastructure for next-generation computational social science, with potential applications in scalable social experimentation and AI-enabled governance platforms.