AgentSociety 2

AgentSociety 2 is a modern, LLM-native agent simulation and research-orchestration platform for computational social science. It provides a unified async framework for agents, environment modules, and research skills, with multiple entry points from the CLI and REST API to integrated research environments.

PyPI Version Python Versions License

Research Background and Platform Positioning

Scientific inquiry has historically expanded the scope, scale, and depth of problems through paradigms such as empirical, theoretical, computational, and data-intensive research. In the digital era, large language models and autonomous agents are reshaping how knowledge is produced and how teams collaborate. For social science, agents can play two complementary roles: as agent scientists that support cognitive work such as literature synthesis, ideation, and experiment orchestration; and as scalable, repeatable silicon subjects for surveys, interviews, and interventions in controlled simulations, enabling counterfactual and mechanistic exploration under cost and ethical constraints.

AgentSociety 2 implements this vision as an end-to-end agent research orchestration framework. Capabilities are organized around four paradigms: empirical experiment design, execution, and analysis; theoretical literature synthesis and hypothesis building (with optional cross-disciplinary workflows); computational agent–environment simulation (extending the AgentSociety 1 tradition with a more modular architecture); and data-intensive acquisition, integration, and inference. Research skills together with the AgentSociety orchestrator form a closed-loop pipeline from literature and hypotheses through computational experiments to analysis and manuscript preparation.

The platform also stresses that meaningful scientific insight still comes from human researchers. The challenge is to reduce engineering burden while keeping human and agent capabilities aligned and collaboration controllable. AgentSociety 2 provides structured workspaces through an IDE-style Integrated Research Environment (IRE) and visualization/extension entry points (backend API, CLI, frontend, and extensions), enabling researchers and multi-agent teams to iterate under transparent control and turn early ideas into reproducible simulations. See Core Concepts for architecture, routing, and storage; see Using Agents and Agent Skills for agents and skills.

Key Features

  • LLM-Powered Agents: Create intelligent agents with personalities, memories, and reasoning capabilities, powered by large language models.

  • Flexible Environment Modules: Build custom simulation environments with composable tools and state management.

  • Async-First Design: High-performance asynchronous architecture for efficient multi-agent simulations.

  • Replay and Analysis: Built-in SQLite-based storage for experiment tracking and analysis.

  • Research Skills: Built-in LLM-native workflows for literature search, hypothesis generation, experiment design, and paper writing.

  • REST API: Independent FastAPI-based backend service for external integrations.

  • CLI Tool: Powerful command-line interface for experiment execution and progress tracking.

  • Extensible: Easily extend with custom agents, environments, and tools.

Installation

pip install agentsociety2

See Installation for detailed installation instructions.

Quick Start

Interact with agents through the async API on AgentSociety (e.g. ask / intervene) using asyncio. See Quick Start for a minimal example.

Documentation

Getting Started:

User Guide:

Developer Guide:

Reference:

Links

Search