The Tsinghua University team has launched AgentSociety², a novel research tool for social science based on large model agents. The system places AI social scientists and silicon-based subjects within the same operational environment, enabling direct integration of literature comprehension, hypothesis generation, experimental design, simulation execution, result analysis, and paper writing. The system supports seven types of experiments covering micro-level behaviors, meso-level network dynamics, and macro-level urban contexts, capable of simulating complex social phenomena such as the emergence of social norms, public goods games, information cocoons, opinion polarization, urban mobility, and disaster response. The breakthrough of AgentSociety² lies in enabling AI to simultaneously fulfill the dual roles of research assistant and experimental participant, transforming social science hypotheses into executable experiments and providing a new infrastructure for computational social science.Author and source: AI New Era
When large language model agents are no longer just chat assistants that answer questions, but begin to possess memory, goals, relationships, and behaviors, a new "scientific imagination" becomes possible: Can we build a simulated "society" within a computer that is experimentable, intervenable, and reproducible? If such a society can be run, observed, and intervened upon, might it open up a new approach to social science research?
In 2025, the team from Tsinghua University took a significant step forward with AgentSociety—a large-scale social simulator built on large language model agents and first principles. AgentSociety integrates language model-driven agents, real-world social environments, and a large-scale simulation engine to generate tens of thousands of agents and simulate millions of agent-to-agent and agent-to-environment interactions, enabling the modeling of complex societal issues such as opinion polarization, information diffusion, universal basic income, hurricane impacts, and urban sustainability.
The key significance of AgentSociety lies in its ability to enable AI agents to enter social simulations at a large scale for the first time, allowing researchers to observe how social behaviors and group dynamics emerge from individual interactions in a computational environment.
Figure 1. AgentSociety² integrates four research paradigms—empirical research, theoretical modeling, computational simulation, and data-intensive methods—through agents to form an "intelligent integration paradigm" for social science research.
Today, the AgentSociety series undergoes a new leap forward, as the Tsinghua University team introduces AgentSociety²: If AgentSociety-1 answered the question of “how to enable AI agents to form a society,” AgentSociety² goes further to address “how to turn this AI society into a genuine laboratory for social science research.”
Paper link: https://agentsociety2.fiblab.net/paper/AgentSociety2.pdf
Official website link: https://agentsociety2.fiblab.net/
Code link: https://github.com/tsinghua-fib-lab/agentsociety/
This is also AgentSociety²’s most fundamental breakthrough: it no longer merely simulates society—it makes social science research itself executable.
AgentSociety² places AI social co-scientists and silicon participants within the same operational environment, enabling literature comprehension, hypothesis generation, experimental design, simulation execution, result analysis, and paper writing to be directly connected to the simulated societies under study.
Figure 2. Development timeline of the large-scale social simulator AgentSociety-1
From "Social Simulation" to "Executable Social Science"
Social sciences have long faced a fundamental challenge: many important questions cannot be repeatedly tested in the real world.
Traditional social sciences rely on surveys, interviews, laboratory experiments, natural experiments, and statistical analysis to understand social phenomena, while computational social science further incorporates large-scale data and simulation models. However, in practical research, theory, data, experiments, simulations, and writing are often scattered across different tools.
After researchers propose a hypothesis, they must manually translate it into experimental protocols, agent configurations, environment rules, intervention plans, and analysis scripts—a process that is both complex and difficult to reuse.
The goal of AgentSociety² is precisely to bridge this gap. It proposes an integrated research environment designed for the social sciences, enabling researchers to start from a high-level question and progressively complete literature review, hypothesis generation, mechanism modeling, experiment configuration, simulation execution, result interpretation, and research presentation.
Unlike general AI scientist systems, AI in the social sciences is not merely an assistant to researchers, because the subjects of social science research are themselves "human beings" and "social processes."
Therefore, the large model agents in AgentSociety² assume two roles: one is AI social scientists, who assist researchers in organizing research workflows; the other is silicon participants, who act as social subjects within configurable social environments, engaging in interactions, responding to interventions, and generating behavioral data.
Figure 3. AgentSociety²: From a simulation-centric silicon-based subject simulation system to a dual-role research ecosystem where silicon-based subjects and silicon-based scientists collaborate.
This dual-role design is the most academically groundbreaking aspect of AgentSociety², as it means AI does not merely assist humans in writing code, searching literature, or generating reports—but simultaneously enters both the “researcher side” and the “subject of study side.”
Researchers pose questions, AI social scientists assist in reviewing literature, formulating hypotheses, designing experiments, and organizing analysis; silicon participants act within simulated societies to generate observable behavioral trajectories and collective outcomes. Both are placed within a single auditable execution environment, enabling social science hypotheses to be translated into agent behaviors, environmental rules, intervention protocols, and measurement metrics.
In other words, AgentSociety² does not have AI directly provide answers to social science questions, but rather helps researchers transform social science problems into executable experiments.
This is the true meaning of "executable social science." In the past, a social science hypothesis was typically written in a paper and relied on subsequent surveys, experiments, or data analysis for validation.
AgentSociety² aims to further transform assumptions into executable experimental structures. For example, “Will recommendation systems reinforce information cocoons?” can be translated into exposure rules and user selection mechanisms within a social media environment; “Why does cooperation decay in public goods games?” can be translated into participant dynamics, payoff structures, punishment mechanisms, and repeated interaction processes; “How do disaster warnings influence crowd movement?” can be translated into event shocks, information dissemination, and behavioral responses within urban spaces. Social science questions thus cease to be merely described—they can now be constructed, executed, intervened upon, and compared.
AI Social Scientist: A Closed Loop from Literature, Hypotheses, Experiments, and Data to Papers
To achieve this goal, AgentSociety² has built an AI social scientist workflow designed for the social sciences.
The system connects multiple stages of social science research into an executable, end-to-end workflow through the harness layer, skill library, sub-agents, tool interfaces, and phased processes—from defining the research topic and conducting literature reviews, to generating hypotheses and designing experiments, followed by simulation execution, result analysis, and paper writing.
Researchers are not excluded from the process but retain control at critical junctures, including hypothesis revision, parameter setting, intervention design improvement, and result interpretation. AgentSociety² thus functions more like an "AI co-pilot for social science": AI handles massive literature processing, workflow organization, simulation execution, and result aggregation, while human researchers focus on posing critical questions, assessing theoretical value, and interpreting mechanistic implications.
Figure 4. Silicon-based social scientists support the entire scientific research process, from defining research topics and generating hypotheses, through experimental design and simulation configuration, to simulation execution, result analysis, and report generation.
Agent-based environment: Transforming social rules into executable experimental scenarios
Another key innovation of AgentSociety² is transforming social environments into callable, composable, and generatable experimental modules. In social sciences, environments are highly complex: public goods games require rules, social media needs recommendation mechanisms, urban mobility demands spatial constraints, disaster response depends on event evolution, and psychological experiments require task workflows.
AgentSociety² encapsulates public goods games, prisoner’s dilemmas, trust games, psychological experiments, social media spaces, event spaces, economic spaces, and mobility spaces as agentic environments, enabling researchers to combine experimental environments according to their specific research questions.
The CodeGenRouter proposed in the system further translates natural language intentions into verifiable environmental operations, allowing researchers to define experimental requirements in a higher-level language rather than writing complex simulation code from scratch, with the system generating executable environment calls.
Figure 5. The agentized experimental environment transforms the agent's natural language intent into executable environmental operations through a unified interface, AST parsing, CodeGenRouter, code caching, and a secure execution mechanism.
Silicon-based subjects: Enabling social experiments with a functional population
In terms of agent design, AgentSociety² has moved from traditional long prompts or fixed workflows to a skill-based architecture.
As the complexity of social experiments increases, character backgrounds, experiment rules, environmental states, tool descriptions, historical interactions, and phase objectives will continue to grow. Including all of them in a single prompt not only increases costs but also risks losing critical constraints.
AgentSociety² decomposes observation, perception, planning, memory, and decision-making rules from specific experiments into reusable skills, enabling agents to dynamically load capabilities relevant to their current tasks.
Meanwhile, each agent has its own independent workspace to store profiles, states, memories, logs, and checkpoints. This means that agents in AgentSociety² are not one-time text generators, but traceable agents that maintain state, update memories, and record trajectories over long-term simulations.
Figure 6. AgentSociety² extends social agents to general silicon-based subject agents, where each agent records its state, memory, and behavioral trajectory in an independent workspace and dynamically invokes skills such as observation, planning, memory, and cognition within the ReAct loop, enabling long-term, traceable, and reproducible social experiments.
Seven categories of experimental validation
From Individual Psychology to Urban Disasters
To demonstrate the system's versatility, the research team designed seven categories of multiscale social science experiments, covering micro-level behavioral experiments, meso-level network dynamics, and macro-level urban contexts.
Figure 7. Demonstration of AgentSociety² experiment cases
At the micro level, AgentSociety² supports the emergence of social norms, public goods games, and psychological surveys to study cooperation, punishment, self-bias, and indirect cognitive processes.
At the meso level, the system simulates information cocoons and opinion polarization, allowing researchers to modify recommendation rules, user choices, and content exposure mechanisms to observe how group fragmentation emerges.
At a macro level, AgentSociety² is applied to urban mobility and disaster response scenarios to simulate crowd behavior during daily travel and under crisis conditions.
This demonstrates that AgentSociety² is not merely a system designed for small-scale dialogue experiments; it can connect individual psychology, group interactions, platform mechanisms, urban behaviors, and public governance issues.
From this perspective, the true significance of AgentSociety² lies in providing a new infrastructure for computational social science: enabling researchers to build populations, design environments, implement interventions, observe emergent behaviors, analyze results, and generate reproducible research evidence—all on a single platform.
Figure 8. Experimental results of the AgentSociety² research case
Social science infrastructure in the AI era
For computational social science, AgentSociety² is significant because it provides a new infrastructure for studying complex social systems.
Computational social science has long sought to understand how macro-level social phenomena emerge from micro-level individual behaviors and interactions—for example, why social norms form, why public cooperation declines, why information cocoons deepen, and why disaster responses exhibit group differences.
Traditional social sciences can observe the real world but struggle to intervene; traditional simulation models can control mechanisms but often simplify human behavior.
The emergence of large model agents has enabled researchers to construct simulated individuals that more closely reflect real-world behavior. AgentSociety² further connects these simulated individuals, social environments, and research workflows, allowing researchers to build populations, design environments, implement interventions, observe emergent phenomena, analyze results, and generate reproducible evidence—all on a single platform.
More importantly, AgentSociety² is not about replacing social scientists with AI, nor does it claim that silicon participants can unconditionally substitute for real humans. Instead, it emphasizes a human-in-the-loop research model, ensuring that human researchers retain judgment and control at critical junctures.
AI social scientists expand the space of explorable mechanisms, reduce engineering burdens, and improve the efficiency of experimental organization; human researchers provide goals, constraints, theoretical judgments, and final interpretations.
This collaborative relationship enables social sciences to move beyond descriptive studies of “what happened” toward mechanistic studies that explore “why it happened” and “what would happen if the mechanisms were changed.”
Figure 9. AgentSociety²: An Integrated Research Environment for Human-AI Collaboration
Today, as AI for Science advances rapidly, AI Scientists have entered fields such as machine learning, biomedicine, chemistry, and materials science. However, the objects of social sciences are more complex, as they study social processes composed of people, relationships, institutions, spaces, and information environments.
AgentSociety² provides an answer tailored to this specialized domain: when an AI Scientist enters the social sciences, it should not merely be an assistant capable of writing papers, but rather an integrated research environment that connects AI social scientists, silicon participants, and agentic environments. It transforms social science from human-managed processes to human-AI collaborative experimentation; from post-hoc analysis of social phenomena to proactive testing of social mechanisms; and from isolated research projects to accumulable, reusable, and auditable research infrastructure.
This is the transformative change brought by AgentSociety²: AI is not just entering society—AI is beginning to help us study society.
In the future, AgentSociety² is expected to be applied to key scenarios such as platform governance, public policy, urban management, disaster response, collective decision-making, social psychology, and AI safety, providing new computational tools and experimental environments for understanding complex social systems.
Its goal is not to replace the real world, but to open up a larger, more controlled, and more reproducible laboratory for the social sciences beyond it.