AgentSchool: An LLM-Powered Multi-Agent Simulation...

AgentSchool: An LLM-Powered Multi-Agent Simulation... | AI Research

Key Takeaways

AgentSchool is an LLM-powered multi-agent simulation platform designed to act as a "wind tunnel" for educational research.
In this work, we introduce AgentSchool, an LLM-driven multi-agent simulator that models learning as state transition rather than prompted behavior.
Further, AgentSchool generates plausible traces of peripheral participation, clique formation, aggressor-induced cohesion, and opinion-leader emergence consistent with classroom social theories.
As large language models are increasingly deployed in classrooms, researchers face a significant challenge: real-world educational trials are slow, ethically sensitive, and difficult to reverse.
AgentSchool addresses this by providing a virtual environment where researchers can model students, teachers, and school scenarios as evolving agents.

Paper AbstractExpand

Despite the rapid deployment of LLMs into classrooms, validating educational AI remains uniquely intractable: interventions act on developing learners whose cognitive and social trajectories are irreversibly shaped, while real-world trials are slow, ethically constrained, and institutionally locked. LLM-based educational simulators have emerged as a potential remedy, but many still collapse learning into persona-conditioned role-play and, when optimized only to reproduce existing classrooms, can structurally penalize the institutional novelty that pedagogical reform requires. In this work, we introduce AgentSchool, an LLM-driven multi-agent simulator that models learning as state transition rather than prompted behavior. AgentSchool couples cognitively growable student agents -- equipped with weighted subject knowledge graphs, thinking-workflow pools, and explicit misconceptions -- with adaptive teacher agents that plan, scaffold, and reflect along the Zone of Proximal Development, embedded in a configurable scenery generator that situates instruction within both formal and informal learning fields, and a multi-scale simulator that decouples interaction scale, temporal granularity, and simulation duration. Experiments show that structured student agents produce more differentiated mastery and misconception traces than a baseline simulator, while teacher-agent comparisons show backbone-dependent patterns consistent with ZPD-informed adaptation. Further, AgentSchool generates plausible traces of peripheral participation, clique formation, aggressor-induced cohesion, and opinion-leader emergence consistent with classroom social theories. Beyond its role as an educational research instrument, AgentSchool frames education as a socially meaningful testbed for long-horizon memory, multi-agent coordination, and future institutional reasoning under organizational pressure.

AgentSchool is an LLM-powered multi-agent simulation platform designed to act as a "wind tunnel" for educational research. As large language models are increasingly deployed in classrooms, researchers face a significant challenge: real-world educational trials are slow, ethically sensitive, and difficult to reverse. AgentSchool addresses this by providing a virtual environment where researchers can model students, teachers, and school scenarios as evolving agents. Instead of relying on simple role-play, the platform simulates the complex, long-term cognitive and social trajectories of learners, allowing for the testing of educational interventions and policies before they are implemented in actual schools.

Modeling Learning as a Process

Unlike traditional simulators that treat learning as a series of prompted responses, AgentSchool models it as a state transition process. Student agents are equipped with "growable" internal states, including weighted subject knowledge graphs, specific thinking workflows, and explicit misconceptions. These students interact with adaptive teacher agents that are designed to plan, scaffold, and reflect based on the "Zone of Proximal Development"—the gap between what a learner can do alone and what they can do with guidance. By tracking these internal states, the system can observe how a student’s understanding changes over time rather than just measuring the correctness of a single answer.

A Configurable Educational Environment

The platform uses a "scenery generator" to create diverse learning fields, ranging from formal classroom instruction to informal social interactions. This allows researchers to study how different environments—such as specific seating arrangements, group dynamics, or technological tools—influence learning outcomes. The simulator is multi-scale, meaning it can decouple the interaction scale, the speed of events, and the total duration of the simulation. This flexibility enables researchers to observe both immediate instructional feedback and the long-term development of student agency and social identity.

Research Insights and Social Dynamics

Experiments conducted with AgentSchool have demonstrated that structured student agents produce more nuanced mastery and misconception patterns than baseline simulators. When observing informal social scenes, the platform successfully generated realistic behaviors consistent with classroom social theories, such as the formation of cliques, the emergence of opinion leaders, and the effects of peer participation. These results suggest that AgentSchool can serve as a valuable testbed for studying complex social phenomena like multi-agent coordination and the long-term impact of institutional pressures on the educational experience.

Future-Oriented Validation

AgentSchool is intended to help researchers move beyond evaluating AI based solely on functional accuracy or user satisfaction. Because educational interventions shape a learner's habits, epistemic trust, and social identity, the authors argue that validation must account for developmental trajectories. By providing a safe, inspectable, and ethically responsible sandbox, AgentSchool allows for "what-if" reasoning about future educational policies and AI-mediated teaching arrangements that do not yet exist, helping to bridge the gap between rapid technological change and slower institutional reform.