From MOOC to MAIC: Reshaping Online Teaching and Learning through LLM-driven Agents

Synthesis: MAIC (Massive AI-empowered Course) proposes a new paradigm for online education that replaces the MOOC's "one video for N students" broadcast with an LLM-driven multi-agent classroom of "N agents for 1 student." By building all agents on a unified Large Language Models (LLMs) foundation, MAIC balances scalability with adaptivity — and collapses course-production cost from roughly $25,000 and 60 hours per MOOC to under $2 and 30 minutes. Piloted at Tsinghua University across two courses with 100,000+ learning records from 500+ students, it deploys specialized Teacher, Assistant, Classmate, and Analyzer agents and is released open source as OpenMAIC.

ArXiv: 2409.03512 Submitted: September 2024 Published version: Journal of Computer Science and Technology (2026) Source code (OpenMAIC): https://github.com/THU-MAIC/OpenMAIC

Overview

MAIC (Massive AI-empowered Course) proposes a new form of online education that addresses the fundamental tension between scalability (MOOC's strength) and adaptivity (MOOC's weakness). Traditional MOOCs serve thousands of learners through one pre-recorded video, struggling to personalize instruction. MAIC replaces this with an LLM-driven multi-agent system that constructs an AI-augmented classroom, shifting from "one video for N students" to "N agents for 1 student".

Key Findings

1. "N agents for 1 student." A suite of specialized LLM-driven agents (Teacher, Assistant, Classmate, Analyzer) constructs an AI-augmented classroom that dynamically adapts teaching to each student's interactions and inquiries, balancing scalability with adaptivity.
2. Course creation at scale. MAIC generates full course materials (slides, textbooks, exercises, videos) from instructor-provided course descriptions — reducing production from ~$25,000 USD and 60 hours per course to under $2 USD and 30 minutes.
3. Unified LLM foundation. Unlike prior systems that used separate models for recommendation, dialogue, and assessment, MAIC builds all agents on a shared LLM foundation, enabling deeper integration across teaching and learning tasks.
4. Standardized course preparation. A Read + Plan workflow transforms static slide decks into highly structured, adaptive learning resources, using Multimodal AI LLMs (e.g., GPT-4V) for slide extraction, description generation, and tree-style knowledge taxonomy construction.
5. Initial pilot evidence. At Tsinghua, two courses ("Toward Artificial General Intelligence" and "How to Study in the University") generated 100,000+ behavioral records from 500+ students over three months, with initial observations suggesting improved engagement versus traditional MOOC formats.
6. Integrated learning analytics. Large-model-powered tools provide quick access to learning data, forecasting of academic outcomes, and automation of interviews and assessments.

Architecture

MAIC deploys a suite of specialized AI agents:

• Teacher Agent: Delivers lectures and core instruction
• Assistant Agent: Provides personalized offline mentoring and exercises
• Classmate Agents: Engage in peer-like dialogue to stimulate discussion and questions
• Analyzer Agent: Diagnoses student performance from quiz results and recommends prerequisite learning paths
• Manager Agent: Maintains order and assists, controlling the class

Pilot at Tsinghua University

• 100,000+ learning records from over 500 students
• Two courses: "Toward Artificial General Intelligence" (TAGI) and "How to Study in the University" (HSU)
• Data drawn from behavioral records, student surveys, and qualitative interviews over a three-month pilot
• Initial observations suggesting improved engagement compared to traditional MOOC formats

Significance

MAIC represents a convergence point for Generative AI, RAG (Retrieval-Augmented Generation), and Agentic AI in education. It demonstrates how LLM-driven multi-agent systems can transform the MOOC paradigm from one-size-fits-all broadcasting to truly adaptive, personalized Intelligent Tutoring at scale. The platform is released as open source under the name OpenMAIC (github.com/THU-MAIC/OpenMAIC), MIT-licensed and actively developed: version 1.0.0 (August 2026) added an agent workbench that plans and revises whole courses, durable sessions that can be canceled, resumed or steered, uploaded materials the agent builds from, and 24 built-in course skills; lessons export as editable .pptx or interactive .html, and the classroom can be self-hosted with Docker or one-click Vercel, or tried on a hosted demo with your own model provider keys. That scale of adoption — more than 38,000 stars by September 2026 — is what makes its "open collaborative hub for AI-driven education" research claim concrete rather than aspirational.

What this means for practice

• Designers. Build every classroom agent on one shared Large Language Models (LLMs) foundation rather than a separate model per function, so teaching, assessment, recommendation, and learning analytics draw on the same representation of the course.
• Instructors. Convert existing slide decks into structured adaptive learning resources with a read-and-plan workflow grounded in a knowledge taxonomy, so generated exercises and explanations stay aligned with the course outline.
• Administrators. Weigh course development decisions against the reported production cost: generating full materials (slides, textbook, exercises, video) fell from roughly $25,000 and 60 hours per course to under $2 and 30 minutes.
• Instructors. Keep humans over generated content and over the class: the pilot routed AI-produced material through subject-matter experts and teaching assistants and retained instructor intervention, and students reported that AI classmates do not replace the discussion and after-class explanation a human teacher provides.
• Researchers. Treat the engagement and thinking gains as preliminary observation and test them against controlled outcome data on completion and self-regulated learning; the open-source OpenMAIC release is positioned as shared infrastructure for that tutoring research.

Limitations

• The Tsinghua pilot covers two courses and 500+ students over three months with no control group; conclusions about engagement rest on initial observation compared with traditional MOOC formats rather than a controlled comparison.
• Higher-order thinking outcomes were measured only as students' perceived impact on pre- and post-course questionnaires (abstract thinking t = 2.32, p = 0.02; critical thinking t = 2.37, p = 0.02), and the authors state that the abilities themselves were not estimated.
• Cost and speed figures for course production come from the team's own generation pipeline in this deployment, not from independent measurement, and the paper notes that some inaccuracies in automatically generated content are expected at scale.

Citation

Yu, J., Zhang, Z., Zhang-li, D., Tu, S., Hao, Z., Li, R., et al. (2024). From MOOC to MAIC: Reshaping Online Teaching and Learning through LLM-driven Agents.