#1020: From Design to Usability: Implementing the Mirror Agents Framework through a Multi-Agent Learn-then-Teach Architecture for Dynamic Metacognitive Calibration

This study advances the Mirror Agents Framework from conceptual design to empirical usability validation in authentic K-12 programming classrooms. We implemented a multi-agent learn-then-teach architecture integrating four coordinated agents and a Teachable Agent that enables students to externalize and calibrate their metacognition through learning-by-teaching cycles. Thirty middle-school students engaged in two iterative learning rounds while the framework implementation captured multimodal behavioral and metacognitive data. Pre-/post-tests (MAI, JOL) and mixed-methods analyses revealed significant gains in metacognitive monitoring accuracy and self-regulation behaviors, supported by qualitative evidence of reduced agent dependence and enhanced strategy transfer. Findings demonstrate the framework embedded system’s usability as a dynamic metacognitive scaffolding tool and highlight the value of integrating multi-agent feedback loops and teachable-agent calibration for fostering autonomous metacognitive development.