AgentSociety²: An Integrated Research Environment Redefining Executable Social Science

From Silicon Participants to AI Social Scientists

AgentSociety² builds on AgentSociety‑1, which in 2025 linked large language model agents, real‑world environments, and a massive simulation engine to study opinion polarization, information diffusion, universal basic income, hurricane impacts, and urban sustainability.

The system introduces a dual‑role architecture: silicon participants act as embodied agents that move, communicate, consume, work, and respond to shocks, while AI Social Scientists orchestrate the entire research workflow—from literature retrieval to hypothesis generation, experiment configuration, result analysis, and manuscript preparation.

AI Social Scientist Workflow

The first core module is a workflow that strings together a harness layer, skill library, sub‑agents, tool interfaces, and staged processes. It translates open research questions into theoretical premises, variable relationships, experimental conditions, interventions, and evaluation metrics, without directly providing answers.

Hypothesis Generation : The system cross‑validates over 20,000 internal top‑journal papers with external sources such as arXiv and OpenAlex, producing structured hypothesis packages that include theoretical basis, experimental design, and simulation modules.

Full‑Process Closed Loop : Semantic research specifications are turned into executable configurations, driving simulations and aggregating numerical results and emergent patterns; during manuscript drafting, AI verifies evidence while humans focus on problem framing and theoretical judgment.

Generative Social Agents (Silicon Participants)

The second core module expands the social agents from city‑level actors to generic, reusable silicon participants suitable for laboratory experiments, psychological surveys, social‑media platforms, and urban environments.

Skill‑Based Architecture : Basic capabilities such as observation, cognition, and planning are encapsulated as reusable skills; agents load needed skills on demand, reducing long‑term simulation overhead and allowing skill combinations to serve as experimental variables.

Independent Workspaces and ReAct Loop : Each agent maintains its own workspace for state, memory, and behavior traces. In a ReAct loop the agent executes real actions and writes back results, making the participant fully traceable, replayable, and comparable across micro‑ to macro‑level analyses.

Agentic Environments

The third core module treats the experimental environment itself as an executable mechanism. Public‑goods games, prisoner's dilemma, social‑media recommendation systems, and urban mobility constraints are packaged as modular environment components.

Modular Environment : Researchers compose scenarios, rules, and interventions to directly translate theoretical hypotheses into runnable experiments.

CodeGenRouter : Natural‑language intents are parsed into abstract syntax trees, cached, and safely executed, cutting large‑model calls by roughly 66.5 % and turning the environment into a composable, auditable mechanism container.

Experimental Validation

To demonstrate capability, the team built experiments at three scales:

Micro‑level: social norm emergence, public‑goods games, and psychological surveys to study cooperation, punishment, self‑evaluation, bias, and cognition.

Mesoscopic: information‑filter bubbles and opinion polarization by adjusting recommendation rules, user choices, exposure, and social structure, observing group fragmentation and testing interventions for diversity.

Macro‑level: city mobility and disaster response where silicon participants navigate spatial constraints, time‑bound events, and crises, simulating daily travel and crowd behavior under emergencies.

These experiments show that AgentSociety² is not limited to small dialogue tasks; it can link individual psychology, group interaction, platform mechanisms, and urban dynamics, enabling researchers to build populations, design environments, apply interventions, observe emergence, analyze outcomes, and archive reproducible evidence.

Re‑organizing the Social Science Research Process

AgentSociety² does not aim for a "fully automatic scientist". Instead, it redesigns the organization of research: humans remain central for problem definition, theoretical judgment, and interpretation, while the agent system handles literature search, hypothesis structuring, environment construction, experiment execution, analysis, and manuscript drafting, forming a traceable, iterative closed‑loop workflow that promotes human‑AI collaboration rather than replacement.

Website: https://agentsociety2.fiblab.net/
Paper: https://agentsociety2.fiblab.net/paper/AgentSociety2.pdf
GitHub: https://github.com/tsinghua-fib-lab/agentsociety/