AgentSociety
Build LLM agent simulations and social science experiments with a self-hosted Python framework that connects to a configured model provider.

AgentSociety is a framework for social scientists studying behavior through LLM-driven simulations. Its AgentSociety 2 package combines social experiments with research workflows.
- Research tasks: Search literature, generate hypotheses, design experiments and write papers.
- Custom simulations: Build agents and modular environments with pluggable tools and different reasoning patterns.
- Social experiments: Study opinion polarization, public-goods games and urban mobility through reusable cases.
- Execution and replay: Run agents through Ray and inspect experiments with replay, stored results and distributed tracing.
- Tool integration: Connect additional tools through the Model Context Protocol (MCP).
The Python framework runs on your own infrastructure. Its model access requires an API key for OpenAI, Anthropic or another LiteLLM-supported provider. Hosted Coder workspaces are a separate offering.
The core uses Apache 2.0. The legacy commercial components are excluded from that open-source license.