中

AI Education Competition Shifts from Models to Agents: A Quiet Paradigm Transfer

AI Agents, AI Education, TAL Education, Capability-Oriented —

At last week’s Harvard China Forum, TAL Education Group President Peng Zhuangzhuang said something that struck me: “The competition in AI education is shifting from the model layer to the agent layer.”

Sounds abstract, but combining recent education industry dynamics, I think I understand what he means.

From “Knowledge-Oriented” to “Capability-Oriented”

Peng made a key point: education’s paradigm is shifting from “knowledge-oriented” to “capability-oriented.”

This sounds like education philosophy cliché, but in the AI context, it means something entirely different. Traditional online education products essentially do “knowledge transfer”—moving textbook content to videos, questions to question banks, making it easier for students to access knowledge.

But in the AI era, knowledge access barriers have dropped extremely low. Want to learn calculus? GPT-4 can explain it. Want to practice speaking? AI can chat with you for 30 minutes. So where’s the value in education products?

Peng’s answer: Shifting from “teaching knowledge” to “building capabilities.”

For example, learning programming:

  • Traditional model: Watch video tutorials → Do exercises → Take tests for verification

  • AI agent model: Give you a real project → AI as mentor guides you from zero to implementation → Develop problem-solving, debugging, documentation-reading capabilities in the process

The latter is closer to “apprenticeship,” with the AI agent playing the “master” role.

Agent vs LLM: Fundamental Differences in Education Scenarios

Many might ask: Can’t GPT-4 be a teacher too? Why need “agents”?

The difference: LLMs are “Q&A-based,” agents are “task-based.”

  • LLM: You ask “How to solve this problem,” it gives you the solution steps. You ask, it answers, interaction ends.

  • Agent: You give it a goal “Help me prepare for finals,” it breaks down tasks, creates plans, supervises your execution, adjusts strategies based on your performance, even proactively reminds you to review.

In education scenarios, the latter is clearly more valuable. Because learning isn’t a one-time Q&A, but a continuous process. Students need not just answers, but guidance, feedback, motivation—exactly what agents excel at.

TAL has already made attempts here. For example, their AI practice companion agent not only grades homework but also proactively pushes targeted exercises based on student weaknesses, even simulating “teacher cold-calling” scenarios to keep students focused.

Education Industry’s “Disintermediation”

Peng also mentioned at the forum: “AI will drive disintermediation in the education industry.”

Harsh words, but I think he’s pointing out a trend: traditional education intermediaries—training institutions, tutoring platforms, even some teachers—if they only provide “knowledge transfer” value, could indeed be replaced by AI.

But this doesn’t mean teachers will lose jobs. Instead, truly valuable teachers will transform from “knowledge transmitters” to “learning guides”—no longer classroom protagonists, but AI agents’ “trainers” and “supervisors.”

I’ve seen cutting-edge edtech companies where teachers already customize AI agent behaviors: setting teaching styles, adjusting difficulty curves, defining feedback mechanisms. Teachers’ work shifts from “lesson prep, lecturing, grading” to “designing learning paths, configuring AI parameters, handling exceptions.”

This is more like being an “education product manager” than a traditional “teacher.”

A Bigger Picture

The shift from model layer to agent layer isn’t just education—it’s a trend across the entire AI application layer.

In early LLM days, people focused on “what models can do”—write code, draw images, write articles. But now, more people focus on “what agents can do”—automatically complete workflows, continuously learn and optimize, collaborate with humans.

The logic: AI’s value isn’t in single-output quality, but in embedding into human daily workflows as genuinely useful “assistants.”

Education is just one reflection of this trend. Healthcare, law, design—all fields requiring professional knowledge + continuous service may experience similar paradigm shifts.

An Open Question

I want to raise a question: In education scenarios, do you prefer AI as a “super teacher” (can teach anything) or a “learning partner” (learns with you)?

My feeling is, currently people want the former—after all, quickly getting answers is a need. But long-term, the latter might be more valuable—because learning isn’t just acquiring knowledge, but developing thinking styles. And an AI partner growing with you might better help build that mindset than an omniscient teacher.

Related Posts