Orchestrated Pedagogical Agents for Efficient Language Learning: An MCP-Enabled Meta-Agent Framework

Abstract

Large language model (LLM)-backed tutors can provide personalized explanations and conversational practice, but current deployments often take one of two suboptimal forms: a monolithic general-purpose agent with inconsistent pedagogy and shallow scaffolding, or a fragmented multi-bot ecosystem that imposes cognitive overhead by forcing learners to switch between specialist chatbots and disrupting learning flow. We introduce OPA (Orchestrated Pedagogical Agents), a meta-agent framework that coordinates task-specific tutors through an explicit pedagogical policy grounded in learning science. OPA leverages the Model Context Protocol (MCP) to standardize access to learner state, practice banks, content repositories, and logging, enabling modular design and reproducible experimentation. OPA implements a progressive tutoring protocol emphasizing minimal effective help, diagnostic learner production, adaptive micro-practice, and scaffolding fading. We outline a planned three-condition evaluation (monolithic, manual multi-bot, and OPA) measuring learning efficiency (gain/min), transfer, delayed retention, and learner experience. This paper presents OPA’s architecture and protocol as a design contribution; empirical validation is reserved for future work.