When LLMs stop chatting and start doing sociology

What it does

AgentSociety 2 is an LLM-native simulation framework built for social science research. It orchestrates agents through multiple reasoning patterns—ReAct, Plan-Execute, Code Generation, and Two-Tier routing—inside modular environments, and records everything to a SQLite replay log. The repository also still carries a legacy v1 package for city-scale urban simulation, though active development is squarely on the research-oriented v2 stack.

The interesting bit

The framework treats the research process itself as an agent skill: built-in capabilities include literature search, hypothesis generation, and experiment design, while paper-writing workflows are available through an external paper-toolkit plugin. It also speaks MCP, so tools can be hot-plugged without rewiring the agent core.

Key highlights

• Designed from scratch around LLM-driven agents rather than retrofitting a traditional ABM engine
• Multiple reasoning routers (ReAct, Plan-Execute, CodeGen, Two-Tier) supported out of the box
• Research-oriented toolset covering literature search, hypothesis generation, and experiment design
• Full SQLite-based experiment replay for reproducibility
• MCP integration for extensible, modular tool use
• Ships with a React frontend and a VS Code extension (ai-social-scientist)

Caveats

• Assumes a remote LLM API key (OpenAI, Anthropic, or litellm-compatible); local or offline operation is not discussed
• Legacy v1 city-simulation packages remain in the monorepo but are explicitly excluded from active CI and security scanning
• End-to-end paper workflows depend on an external paper-toolkit plugin, not the core library

Verdict

Social scientists and AI researchers who need reproducible, structured LLM experiments should look here. If you want lightweight agent orchestration without the research scaffolding, it may be overbuilt.