A Prompt-Driven Modular Framework for LLM-Based Agents in Scalable Interactive Learning Systems

Abstract

This paper presents a prompt-driven modular architecture for integrating Large Language Model (LLM)-powered agents into adaptive learning systems. The framework enables real-Time, context-Aware dialogue and feedback through natural language configuration, eliminating the need for retraining or code changes. It features a four-layer design-covering environment control, agent logic, UI/UX, and feedback-implemented in a Unity-based learning application. A case study in medical education demonstrates how structured prompt templates support agent role assignment, behavior definition, and fallback strategies. Evaluation includes prompt reusability, dynamic role switching, and LLM accuracy comparison across multiple models, which show high correctness on domain-specific tasks. The system also supports low-cost deployment, making it feasible for scalable and accessible intelligent learning environments. This work offers a generalizable and configurable agent design method, contributing to the development of LLMintegrated tutoring systems and broader applications in natural language-driven interactive learning.