From Chatbot to Co-Worker: What "Agentic AI" Actually Means for Teachers

The AI most teachers know

If you've been anywhere near teacher Twitter or the news lately, you'll have seen the term agentic AI cropping up. It sounds like another bit of tech jargon designed to make you feel behind.

But it does describe something genuinely different from what most teachers are currently doing with AI. It's worth understanding because it signals where things are heading.

By "agentic AI", I mean systems made up of software "agents" that can plan and act with some autonomy to achieve a goal, rather than just generating text in response to a prompt.
Source: AI Insights: Agentic AI (GOV.UK)

Most teachers are familiar with AI in its simplest form:

You go to ChatGPT, you type a question, you get a response. Maybe you ask it to explain quantitative easing for Year 12, or to suggest some discussion questions on Macbeth. It talks, you read, you decide what to do with it.

That's AI as an assistant that mostly talks back rather than does work. And I don't mean that dismissively. It can be genuinely useful. But the relationship is simple: you ask, it answers. You still do the thinking about what to ask, how to use the output, and whether it's any good.

A lot of teachers I speak to are still here. And if that's you, there's absolutely nothing wrong with that.

The next step most teachers take

Some teachers have moved beyond asking AI questions and started asking it to make things.

• Write me a worksheet.
• Draft some feedback comments.
• Create a starter on climate change.

This is a step forward because you get classroom materials, not just explanations. But you're still the project manager.

You're crafting the prompt, checking whether the output matches your spec, copying it into a document, reformatting, realising it doesn't quite align with the mark scheme, re-prompting, and gradually assembling something usable from a few attempts.

You're still doing a lot of the work. The AI is still doing a lot of the talking.

What's actually different about "agentic" AI

Here's where the shift happens.

Agentic AI doesn't just respond to you. It makes decisions about how to complete a task.

A helpful way to think about it is the difference between asking someone a question and delegating a task.

When you ask a question, you get an answer and then you decide what to do next.

When you delegate, you say "here's what I need" and the other person figures out the steps, gathers what they need, makes judgement calls along the way, and comes back with a finished output.

That's what "agentic" means. The AI isn't waiting for you to tell it what to do at each step. You define the goal, it works out the "how".

What this looks like in practice

I'll use our Question Generator at TeachEdge as a concrete example, because it shows the difference clearly.

When a teacher uses the tool, they select a subject, an exam board, a topic, and a question type. That's it. That's the delegation.

Behind the scenes, something different from a chatbot conversation is happening. The system receives the brief and starts reasoning about what it needs.

If the question type calls for a real-world extract, for example data on UK inflation for an Economics extract question, the AI can decide it needs current information and then use search as a tool to gather it.

The key point is that it's not "one search and hope for the best". It can iterate:

• search
• evaluate what it found against constraints we set (source quality, relevance, freshness)
• try again if needed, within strict limits

The aim is to reduce the "teacher has to prompt-hack it into behaving" problem, and replace it with a structured workflow.

Once it has sufficient source material, it synthesises everything into a single coherent output:

• an exam-appropriate extract grounded in real data
• a question that genuinely requires the extract to answer
• a mark scheme aligned to that exam board's assessment objectives
• the associated metadata (command words, topic area tags, difficulty, estimated time)

The important bit is that the extract, question, and mark scheme are generated as one connected task, with consistency checks along the way, rather than me stitching three separate chatbot outputs together.

So the teacher gets back a complete, exam-ready question. Their role shifts from project manager to quality controller.

Not "is this formatted correctly?" but "is this the right question for my students right now?"

This is the direction, not the destination

Our Question Generator is one example, but the same pattern is appearing elsewhere.

Anthropic recently launched Claude Cowork, positioned as an agentic "do the task for you" experience aimed at multi-step workflows like organising files and producing formatted documents.
Source: Getting started with Cowork (Claude Help Centre)

It's the same principle: you define the outcome, the system handles the execution.

The direction of travel is clear. AI is moving from telling you things to doing things. From answering your questions to completing your tasks. From chatbot to co-worker.

What this means for teachers

If AI stays as a chatbot, the key skill is prompt engineering: learning to ask the right questions in the right way to get useful outputs.

That's a real skill, but it's mainly about becoming a better manager of a slightly unpredictable tool.

If AI becomes more agentic (and it is), the key skill is something teachers already have: professional judgement.

Knowing what good looks like.

• Is this extract appropriate for my class?
• Does this question actually assess what it claims to assess?
• Does this mark scheme reflect how we'd mark under real exam conditions?
• Is this the right task at the right time for these students, this week?

That's reassuring, I think. The direction of travel isn't towards teachers needing to become more technical. It's towards tools that handle more of the operational legwork, freeing teachers to do what they're trained for: making expert decisions about teaching and learning.

The honest bit

I don't want to oversell this.

Agentic AI isn't infallible. Our Question Generator sometimes picks sources that aren't quite right, or produces questions that need tweaking. Teacher oversight isn't optional. It's the whole point.

If anything, agentic systems make the need for guardrails more obvious. When an AI can take actions (search, pull information, organise files), you want clear boundaries around what it can access, what it's allowed to do, and how you review the outcome.

Anthropic's write-up of "computer use" makes the same point: useful capability, but safety and oversight matter.
Source: Developing a computer use model (Anthropic)

But the nature of oversight is changing.

• Less "fix the AI's output"
• More "evaluate and select"

Less time wrestling with prompts and reformatting, more time thinking about your students.

That's a shift worth understanding. Not because you need to rush to adopt every new tool, but because when you do evaluate AI tools for your classroom, you'll know what to look for:

• Does this tool just talk to me, or does it actually do something?
• Am I still the project manager, or can I be the professional?
• What guardrails and review steps are built in?

The tools that answer those questions well are the ones worth your time.

---

Gary is the founder of TeachEdge and a practising Economics teacher. TeachEdge uses AI to help teachers mark essays and generate exam-ready questions, with teacher oversight at every stage. If you'd like to try the Question Generator, you can find it at teachedge.ai.