School-of-AI: The Multi-Agent Classroom: Inside OpenMAIC and the School of AI

The Simulated Study Group

The standard interface for generative AI in education is a solitary chat window. A user asks a question, and a single model responds. OpenMAIC fundamentally rejects this premise. Instead of treating the large language model as an omniscient tutor, it treats multiple LLMs as a cast of characters.

Built as a Next.js 16 application, OpenMAIC simulates a social learning environment. It provisions distinct agents with specific roles. A lecturer agent delivers the core material. A debater agent challenges the user's assumptions. A grader evaluates the responses in real time. These agents interact not just with the user, but with each other.

Orchestrating the Agents

Running multiple AI agents simultaneously risks creating a chaotic system where models talk over each other or hallucinate the user's intent. OpenMAIC solves this by utilizing LangGraph as its orchestration layer. LangGraph operates as a state machine, meticulously managing turn-taking and context passing.

When a user submits a question, the input does not go directly to an LLM. It hits the orchestrator. The orchestrator evaluates the current state of the classroom and routes the packet to the appropriate agent. Once that agent completes its task, it passes a structured payload back to the orchestrator, which then updates the global context and decides the next step.

Generating the Curriculum

The output layer of OpenMAIC is equally dynamic. The system ingests a plain text document or topic and generates a complete curriculum on the fly. It does not simply return markdown text. It produces slide decks with voice narration, interactive quizzes, and standalone HTML simulations that render natively in the browser.

This allows the multi-agent system to utilize visual aids. A lecturer agent can trigger the generation of a chart, while a debater agent points to a specific data point within that newly generated graphic.

From Numpy to LangGraph

To understand the philosophy behind OpenMAIC, one must look at its origins. The project traces its roots back to the 2018 Avik-Jain/School-of-AI repository. Long before LLM orchestration, this repository served as a curriculum for community-led machine learning workshops.

The 2018 codebase focused heavily on demystifying black boxes. Instead of using high-level libraries like Scikit-Learn to perform facial recognition, the curriculum taught students to implement K-Nearest Neighbors from scratch. It forced learners to flatten 3D face data into 1D vectors manually.

# Legacy 2018 implementation reshaping face data manually
f_01 = np.load('./face_01.npy').reshape(20, 50*50*3)

That same ethos of transparency drives OpenMAIC today. By using an open-source LangGraph architecture, the platform exposes exactly how the AI reasoning process works. The student is not just learning the curriculum; they are interacting with a visible, inspectable cognitive engine.

The Single-Agent vs. Multi-Agent Paradigm

The educational technology landscape is crowded. OpenMAIC differentiates itself by focusing purely on the generative, multi-agent experience. It contrasts sharply with platforms that rely on static content or single-threaded conversations.