AgentSchool: An LLM-Powered Multi-Agent Simulation for Education

Researchers introduce AgentSchool, an LLM-powered multi-agent simulator that models student learning through state transitions rather than simple role-play, featuring cognitively growable student agents with knowledge graphs and adaptive teachers operating within the Zone of Proximal Development. The system addresses the challenge of validating educational AI interventions in real classrooms by creating a configurable simulation environment that reproduces plausible learning outcomes and social dynamics without requiring institutional constraints or ethical compromises of live trials.

AgentSchool represents a meaningful advancement in educational AI validation by moving beyond prompt-based persona simulation toward mechanistic modeling of learning processes. Traditional classroom simulations often fail to capture genuine cognitive development or support pedagogical innovation, instead merely reproducing existing institutional structures. This work decouples learning mechanics from behavioral mimicry by equipping student agents with explicit knowledge representations, misconception modeling, and adaptive learning pathways.

The research addresses a fundamental problem in educational technology: validating interventions on developing learners is ethically fraught and institutionally slow. Real-world trials risk irreversibly shaping cognitive trajectories while remaining subject to regulatory constraints and institutional gatekeeping. Simulation-based testing offers an alternative, though previous approaches collapsed learning into superficial role-play. AgentSchool's multi-scale simulator architecture—separating interaction scale, temporal granularity, and simulation duration—enables researchers to explore pedagogical reform without institutional lockdown.

The technical contribution focuses on structured agent design: weighted knowledge graphs, thinking-workflow pools, explicit misconception tracking, and teacher agents that scaffold instruction along the Zone of Proximal Development. Experiments demonstrate that structured agents produce more differentiated learning traces than baseline simulators, while social dynamics simulations reveal emergent phenomena like clique formation and opinion-leader emergence consistent with established classroom theories.

For the AI research community, AgentSchool frames education as a proving ground for long-horizon memory, multi-agent coordination, and institutional reasoning under organizational pressure. This positions education not merely as an application domain but as a complex testbed for advanced AI capabilities. Future developments will likely focus on validating whether simulated pedagogical interventions transfer to real classroom effectiveness.

• →AgentSchool uses state-transition modeling rather than role-play to simulate authentic student learning with cognitive growth mechanisms.
• →The simulator couples adaptive teacher agents with structured student agents featuring knowledge graphs, thinking workflows, and explicit misconceptions.
• →Experiments show structured agents produce more differentiated mastery and misconception traces than baseline simulators.
• →The system generates plausible social dynamics including clique formation, cohesion patterns, and opinion-leader emergence.
• →AgentSchool addresses the ethical and institutional constraints of validating educational AI through real-world classroom trials.