Is Agentic AI Coming to K-12 Education?

Artificial intelligence has migrated into the work of schools in myriad ways.

Teachers are using it to grade short-answer assessments and to analyze data. Parents are turning to chatbots to get answers to pressing questions about school scheduling or to help their children with homework. Students are using it for help with difficult concepts – with or without schools’ knowledge.

Most of these tools, though, are task-specific. They’ve been coded with sets of rules to enable them to analyze patterns — and not much more.

Yet there’s another form of the technology that can take initiative to make decisions independently, set goals, and respond dynamically to changes without as much human control: agentic AI.

Agentic AI is generally defined as a form of the technology that can operate autonomously to achieve different objectives, with minimal human supervision.

Whereas the artificial intelligence that districts are currently using may automatically grade assignments or flag which students may be at risk, agentic AI could go the extra step to change strategies mid-lesson according to what students need. Or, the tool could be used to diagnose and analyze student struggles and schedule a meeting with an instructional coach, a counselor, or another adult, without a human triggering the action.

Some ed-tech developers believe agentic AI could offer a path for schools to “personalize” instruction and interventions in ways that have eluded educators for decades.

However, school districts are far from making agentic AI a part of their normal operations. Many of the concerns that educators and others have about AI, overall – around data privacy, accuracy, and bias, for instance – also apply to agentic forms of the technology.

Ed-tech companies need to understand the implications of agentic AI as it evolves, said Charles Elliott, head of industry architects for Google Public Sector’s Rapid Innovation Team.

EdWeek Market Brief recently spoke with Elliott at the ASU+GSV Summit — an annual gathering of ed-tech leaders, investors, and policymakers – about agentic AI as K-12 begins to see the early applications of it and what it means for vendors going forward.

Let’s start with the basics: What’s the difference between agentic AI and the AI that we’re already familiar with?

Let’s imagine two worlds. The first one is the more linear, narrow AI, like a chatbot – so a generalist agent that understands a lot of stuff. Maybe it’s even provided some content, like a syllabus. That experience is pretty narrow.

But in the world of an agentic AI system, you perhaps have agents that [say], “I’m going to block 15 minutes for you to read a little bit more about this thing before you get in the class. And by the way, I’m going to make sure you connect with this person ahead of time.”

So it’s a little bit of the assistant on your phone coming to life, but with that bigger, deeper knowledge about the goals you need to achieve.

What role do you see agentic AI playing in transforming education?

[Agentic AI] can be so powerful in education because it’s really good at connecting dots because they’re training on so much data. The benefit there is that it can draw correlations.

What kind of correlations could agentic AI help with in a school setting?

Let’s say I just signed up for an [Advanced Placement] course in high school. It’s my first one, and I don’t know much about it other than, maybe it’s harder than classes I took before. [In this scenario] students are almost briefed by the AI – giving them a little bit more perspective.

So if you’re a student, and you’re nervous about taking your first AP Bio course, the agent can give you some confidence and can point to specific areas in the curriculum. It builds confidence so students are more prepped at that point to ask questions in a classroom setting, and it helps them engage with the content a little bit more directly.

It can prep you by asking questions and then verifying those things, and maybe even at some point, providing some of those preliminary questions back to the instructor before you get there. It’s about how many other questions can [the agent] ask to get that person excited, and how many questions does the agent need to ask of the content that it has access to.

So agentic AI allows for the technology to have increased autonomy and decision-making. How do you mitigate for the increased risk that comes with that?

Agentic AI, like most successful AI systems today, absolutely require humans in the loop. It’s [humans’] expertise, the empathy – all the things that AI is not really capable of. Agentive AI is meant to be assistive. It’s not there to sit in the class with a student. It’s just more about helping them engage.

Ideating with [the technology] is probably one of the best places to start, but you don’t have to use any of the content. It can just help you think differently about the problem or about the cohort of students that [you’re serving].

What are good examples you’ve seen of how agentic AI can be applied to K-12?

The best example I’ve seen comes down to things like grading because it leverages the parts of the models that are pretty well-understood at this point – things like understanding what’s in a document and fitting into a workflow. And it still keeps the human in the loop.

You don’t necessarily always know how many AI agents are being used behind the scenes. But with grading, you can immediately see the efficiency in helping the student by coming up with recommendations once the grading is done.

I’ve also seen in a classroom setting, with the kinds of questions [specific students] are asking, [teachers] might quickly realize, “I should have included some diagram here.” So that level of personalization could be at the one-on-one level, where the [AI] model says, “I’m going to help you with your homework, and I’m going to include more diagrams.”

[Models] can also generate, in addition to the text based-content, things like podcasts for the students, and create these engaging experiences.

What would be your message to vendors about the promise, and limitations, of agentic forms of AI?

Agentic AI does enable a certain level of creativity, whether you’re building an app or trying to build new experiences. But ultimately, being knowledgeable about the types of systems and how they integrate with what you’re building is super important.

Paying attention to things like benchmarks is still valuable because the benchmarks might tell you that you need to pivot to another model. Or there might be some new standards out there for how models interact with certain types of data. It’s all moving really quickly.

Any words of caution you would offer education companies about the technology?

You also want to make sure that you’re not just building confidence, but that you’re actually focusing on the pedagogical principles. You’re giving the model access to materials, and instead of it being super goal-oriented and getting the answer solved as quickly as possible, [you have to] focus on pedagogy, and you can certainly infuse that kind of training into the model.

When we define student success, it’s about skills acquisition, not just giving them answers, and the guardrails are really important to make sure they’re focused on the task at hand while still stimulating that curiosity.