AI Agents in Education: Khan Academy and Duolingo Deploy Autonomous Tutoring Agents to 50M Students

Personalized Tutoring at Unprecedented Scale

The promise of AI in education has long been "a personal tutor for every student." In March 2026, that promise is materializing at scale. Khan Academy and Duolingo — two of the world's largest education platforms — have independently deployed autonomous AI tutoring agents that collectively serve over 50 million active learners, fundamentally changing how personalized instruction is delivered.

Khan Academy's "Khanmigo Tutor Agent," an evolution of its earlier Khanmigo chatbot, and Duolingo's "Max Tutor," an expansion of its Duolingo Max subscription tier, both represent a shift from reactive AI assistants to proactive agentic tutors. These systems do not wait for students to ask questions. They observe, adapt, intervene, and design learning paths autonomously.

Khan Academy's Khanmigo Tutor Agent

Khan Academy's system, launched to all free-tier users on March 1, 2026, after a year-long pilot with 2 million students, is built on Anthropic's Claude 3.5 Sonnet and operates as a multi-agent architecture.

flowchart LR
    CALLER(["Student or Parent"])
    subgraph TEL["Telephony"]
        SIP["Twilio SIP and PSTN"]
    end
    subgraph BRAIN["Education AI Agent"]
        STT["Streaming STT<br/>Deepgram or Whisper"]
        NLU{"Intent and<br/>Entity Extraction"}
        TOOLS["Tool Calls"]
        TTS["Streaming TTS<br/>ElevenLabs or Rime"]
    end
    subgraph DATA["Live Data Plane"]
        CRM[("CRM and Notes")]
        CAL[("Calendar and<br/>Schedule")]
        KB[("Knowledge Base<br/>and Policies")]
    end
    subgraph OUT["Outcomes"]
        O1(["Enrollment captured"])
        O2(["Tour scheduled"])
        O3(["Counselor callback"])
    end
    CALLER --> SIP --> STT --> NLU
    NLU -->|Lookup| TOOLS
    TOOLS <--> CRM
    TOOLS <--> CAL
    TOOLS <--> KB
    NLU --> TTS --> SIP --> CALLER
    NLU -->|Resolved| O1
    NLU -->|Schedule| O2
    NLU -->|Escalate| O3
    style CALLER fill:#f1f5f9,stroke:#64748b,color:#0f172a
    style NLU fill:#4f46e5,stroke:#4338ca,color:#fff
    style O1 fill:#059669,stroke:#047857,color:#fff
    style O2 fill:#0ea5e9,stroke:#0369a1,color:#fff
    style O3 fill:#f59e0b,stroke:#d97706,color:#1f2937

The Diagnostic Agent continuously assesses student understanding through a combination of exercise performance, response timing patterns, and natural language interactions. Unlike traditional adaptive learning systems that rely solely on right-or-wrong scoring, the diagnostic agent analyzes the reasoning process itself. If a student consistently gets algebra problems correct but takes significantly longer on word problems, the agent detects a reading comprehension gap masquerading as a math difficulty.

The Curriculum Agent dynamically resequences learning content based on diagnostic outputs. Rather than following a fixed skill tree, this agent constructs personalized learning paths that adapt in real-time. If a student is struggling with fractions, the agent might route them through a visual geometry unit first, building spatial intuition before returning to numerical fraction operations.

The Instruction Agent handles direct tutoring interactions. It uses the Socratic method by default — asking guiding questions rather than providing answers — but adapts its teaching style based on student engagement signals. Some students respond better to worked examples. Others prefer analogies. The agent tracks which pedagogical approaches produce the best outcomes for each individual learner and adjusts accordingly.

"The pilot results exceeded our expectations," said Sal Khan during an education technology conference keynote. "Students using the Tutor Agent showed a 32% improvement in learning velocity compared to standard Khan Academy exercises, and a 41% improvement compared to passive video watching. But the number that matters most to me is engagement — daily active usage increased by 28% among pilot participants."

Making It Free

The most significant aspect of Khan Academy's deployment is its pricing: free for all users. Khan Academy, as a nonprofit, secured a $25 million grant from the Gates Foundation specifically to fund the compute costs of running the tutoring agent at scale. Additional funding from Google.org and the Valhalla Foundation covers operational costs through 2028.

"The whole point of Khan Academy is free, world-class education for anyone, anywhere," Khan emphasized. "An AI tutor that only works for people who can afford a subscription is not aligned with our mission."

Duolingo's Max Tutor

Duolingo's approach, available to its 8 million Max subscribers, takes a different architectural direction. Built on OpenAI's GPT-4o, Duolingo Max Tutor operates as a persistent conversational agent that maintains a long-term model of each learner's language abilities.

The system introduces several agentic capabilities beyond Duolingo's traditional gamified lesson format.

Immersive Scenario Agent: Instead of isolated vocabulary drills, the agent creates context-rich scenarios — ordering at a restaurant in Paris, negotiating a price at a market in Mexico City, asking for directions in Tokyo. The scenarios adapt to the learner's proficiency level and interests, drawn from a profile the agent builds over time.

Pronunciation Coach Agent: Using Duolingo's speech recognition technology, this agent provides real-time pronunciation feedback during conversational practice. It identifies specific phonemes the learner struggles with and designs targeted micro-exercises. For Mandarin learners, for example, it can detect and correct tone errors on individual syllables.

Cultural Context Agent: Language learning is inseparable from cultural understanding. This agent weaves cultural context into lessons — explaining when to use formal versus informal address in Japanese, why certain phrases are considered rude in Brazilian Portuguese, or how idioms differ between European and Latin American Spanish.

Progress Synthesis Agent: Weekly, this agent generates a comprehensive learning report that summarizes the student's progress, identifies areas for focus, and adjusts the upcoming week's lesson plan. These reports are also shared (with user permission) with human language teachers for students enrolled in blended learning programs.

"Our internal data shows that Max Tutor users are reaching conversational proficiency 2.3 times faster than users on our standard tier," said Luis von Ahn, CEO of Duolingo. "The agent does not just teach vocabulary. It teaches communication."

Learning Outcomes Data

Both platforms have released preliminary learning outcomes data that has caught the attention of education researchers.

Khan Academy's pilot study, conducted with researchers from Stanford's Graduate School of Education, tracked 180,000 students across 12 months. Key findings include a 32% increase in concept mastery speed, a 24% reduction in time spent on already-mastered concepts (the agent avoided redundant review), and a 45% improvement in long-term retention measured at 90-day follow-up.

Duolingo's data, from a randomized controlled trial with 50,000 participants published as a preprint on SSRN, showed that Max Tutor users scored 38% higher on standardized language proficiency tests (CEFR-aligned assessments) after six months compared to a control group using standard Duolingo lessons.

Dr. Rose Luckin, a professor of learner-centered design at University College London, called the results "the strongest evidence we have seen that AI tutoring can meaningfully improve learning outcomes at scale. The key innovation is not the language model itself — it is the agentic architecture that enables continuous adaptation."

Concerns and Challenges

Not everyone is celebrating. The National Education Association (NEA) issued a statement cautioning against over-reliance on AI tutoring, noting that "education is fundamentally a human endeavor that requires empathy, mentorship, and social connection that AI cannot replicate."

There are also equity concerns. While Khan Academy's offering is free, effective use still requires a smartphone or computer and internet access — barriers that affect the very populations that could benefit most from personalized tutoring. Duolingo's Max Tutor is behind a paywall at $29.99 per month, raising questions about whether the best AI tutoring will be reserved for those who can afford premium subscriptions.

Data privacy for minors is another pressing concern. Both platforms comply with COPPA and FERPA regulations, but advocacy groups like the Electronic Frontier Foundation have called for more transparency about how student interaction data is used to train and improve the AI systems.

The Broader EdTech Landscape

Khan Academy and Duolingo are the highest-profile deployments, but they are far from alone. Chegg launched an AI tutoring agent in January 2026 for its 4 million subscribers. Coursera is piloting agent-based mentors for its professional certificate programs. And China's Squirrel AI, one of the earliest adaptive learning companies, deployed a new agentic architecture across 3,000 learning centers in February 2026.

The global AI in education market is projected to reach $32.7 billion by 2028, according to HolonIQ, up from $6.1 billion in 2024. Venture capital investment in EdTech AI companies reached $4.2 billion in 2025, with tutoring agents representing the single largest investment category.

The consensus among education technology analysts is that agentic AI tutors will not replace human teachers but will fundamentally change what human teachers spend their time doing — shifting from content delivery to mentorship, motivation, and the social-emotional dimensions of education that AI cannot address.

Sources

• Khan Academy Blog — "Khanmigo Tutor Agent: Free AI Tutoring for Everyone" (March 2026)
• Duolingo Engineering Blog — "Building Max Tutor: An Agentic Architecture for Language Learning" (March 2026)
• Stanford Graduate School of Education — "Evaluating AI Tutoring at Scale: The Khanmigo Pilot Study" (February 2026)
• HolonIQ — "Global EdTech AI Market Report 2026" (January 2026)
• The New York Times — "AI Tutors Go Mainstream: Can They Deliver on the Promise?" (March 2026)