DeepTutor: Agents Learned to Teach
HKU Data Intelligence Lab shipped DeepTutor v1.0.0 this week, and the numbers speak for themselves -- 12,000 GitHub stars in 39 days, currently trending at 339 stars per day. That is not typical for an education tool.
DeepTutor calls itself agent-native, and for once the label is earned. It is not a chatbot with a study guide bolted on. The architecture is built around TutorBots -- persistent, autonomous agents with independent workspaces and personalities that live across Telegram, Discord, Slack, and other channels. They have a heartbeat system that proactively sends reminders and check-ins. They remember your learning progress, study history, and preferences across sessions.
The v1.0.0 release (beta.2 dropped April 7) rewrote the entire architecture around an agent-native plugin model with CLI and SDK entry points. Five modes share one conversation thread -- Chat, Deep Solve, Quiz Generation, Deep Research, and Math Animator -- with seamless context switching. There is a built-in AI co-writer, RAG-ready knowledge bases from PDFs and Markdown, and guided learning that transforms any material into structured visual learning journeys.
What makes this different from Khan Academy AI or Duolingo chatbot is the agent-first design. The TutorBot is not an add-on -- it IS the product. It runs autonomously, maintains state, uses tools (code execution, web search, paper search), and operates across multiple channels simultaneously. The dual output mode (human-readable and JSON) means other agents can consume DeepTutor output natively.
This is what agents for education actually looks like when you start from the agent and build toward education, rather than the other way around.
https://github.com/HKUDS/DeepTutor
← Back to all articles
DeepTutor calls itself agent-native, and for once the label is earned. It is not a chatbot with a study guide bolted on. The architecture is built around TutorBots -- persistent, autonomous agents with independent workspaces and personalities that live across Telegram, Discord, Slack, and other channels. They have a heartbeat system that proactively sends reminders and check-ins. They remember your learning progress, study history, and preferences across sessions.
The v1.0.0 release (beta.2 dropped April 7) rewrote the entire architecture around an agent-native plugin model with CLI and SDK entry points. Five modes share one conversation thread -- Chat, Deep Solve, Quiz Generation, Deep Research, and Math Animator -- with seamless context switching. There is a built-in AI co-writer, RAG-ready knowledge bases from PDFs and Markdown, and guided learning that transforms any material into structured visual learning journeys.
What makes this different from Khan Academy AI or Duolingo chatbot is the agent-first design. The TutorBot is not an add-on -- it IS the product. It runs autonomously, maintains state, uses tools (code execution, web search, paper search), and operates across multiple channels simultaneously. The dual output mode (human-readable and JSON) means other agents can consume DeepTutor output natively.
This is what agents for education actually looks like when you start from the agent and build toward education, rather than the other way around.
https://github.com/HKUDS/DeepTutor
Comments