On September 8, in Paris, at UNESCO Headquarters, Professor Zhang Yu and Assistant Professor Yu Jifan from Tsinghua University’s School of Education spent 90 minutes introducing OpenMAIC to teachers, principals, and policymakers from different countries and regions during an official event at Digital Learning Week 2026. Participants also submitted their own teaching topics to the Agent and generated courses themselves. This was the first Chinese education Agent team to appear on the UN stage.
For the past two years, the AI industry has been searching for the next high-value entry point after coding. Office software has competed for documents, spreadsheets, and calendars; smart hardware has competed for real-world data; education has long been undervalued. OECD TALIS 2024 shows that about one-third of lower secondary teachers already use AI in their work, and among users, 64% use it to generate lesson plans or classroom activities.
Education scenarios and software engineering share a similar skeleton: define learning objectives, retrieve and verify materials, organize the sequence of knowledge, then generate explanations, slides, interactions, and feedback, while continuously calling models through multiple rounds of revision. OpenMAIC originated from the MAIC project launched in 2024 and was open-sourced in March 2026. As of August 20, the platform had more than 550,000 creators, cumulative interactions exceeding 34 million, and coverage across 154 countries and regions.
The change in magnitude is even more telling. The older version consumed an average of about 358,000 tokens to generate one interactive classroom, while v1.0’s professional workbench consumes an average of about 3.625 million tokens to complete a full creation task—about 10.1 times more than before—with an average of 74.3 model calls behind it. These numbers are not waste; they are the direct result of a longer task chain. AI is shifting from “generating a piece of content” to “undertaking an entire production workflow.”
When AI begins to take on planning, retrieval, generation, revision, and reuse, education becomes a typical long-chain agent task. This aligns with StarWar Technology’s emphasis in its OPC agent collaboration platform and AI training platform: enabling multiple models, tools, and agents to collaborate around a learning goal while making the process manageable and reusable. Education is both a proving ground for agent capabilities and a training ground for teaching people how to collaborate with AI.
Another experiment from OpenMAIC is packaging methods such as the Feynman Technique and spiral curriculum as Skills, and connecting them to workbenches such as DeepSeek Harness and Codex through a standardized SDK. This means teaching methods can be described, invoked, and transferred. Educational capabilities now have the opportunity to be embedded into scenarios built by different developers, much like search, writing, and programming tools.
As foundation models gradually converge, competition among education Agents will shift from single-point features to workflows and ecosystems. Whoever can make more teaching methods reusable and onboard more developers is more likely to define new production standards for educational content. Walking onto the UN stage, OpenMAIC is exporting not just a product, but a methodology for building AI-native learning environments.