Most of what schools and learning teams adopted in the last two years was a chat box. You ask, it answers, the conversation ends. That is useful, but it is not what people mean when they say agentic AI.
An agent is given a goal rather than a question. To reach it, the agent breaks the goal into steps, calls tools to get things done, looks at what happened, and decides what to do next. It keeps going until the goal is met, it gets stuck, or a person tells it to stop.
Chatbot vs. agent: the practical difference
Take a learner who is three weeks behind in a data-analysis course.
- A chatbot waits for the learner to ask something. If they never ask, nothing happens.
- An agent notices the missed submissions in the LMS, reads which topics they covered, drafts a catch-up plan that fits the time left before the deadline, messages the learner, and flags the case to a mentor if there is no reply in two days.
Same underlying model. The difference is the loop around it: a goal, access to tools, memory of what it has already done, and permission to act.
The four parts of an agent
- Goal — a clear outcome, such as "every learner who is behind has a realistic catch-up plan".
- Tools — the actions it is allowed to take: read the gradebook, look up course material, send a message, book a calendar slot.
- Memory — what it has already tried with this learner, so it does not repeat itself.
- Checks — rules it must verify before acting, and the points where a human has to approve.
If any of those four is vague, the agent will be too.
Where agents genuinely help in learning
The best early uses are the ones where the work is repetitive, the rules are clear and the cost of a mistake is small:
- Spotting learners who are drifting before the drop-out point, not after it.
- Turning a learner's wrong answers into a personalised revision set for the week.
- Preparing first-draft feedback on assignments against a rubric, for a mentor to review.
- Handling scheduling, reminders and resource links so mentors spend their time teaching.
Where they should not be left alone
Agents take actions, which means their mistakes are actions too. Anything that affects a learner's grade, record or wellbeing needs a person in the loop. A sensible default is simple: the agent can draft anything, but it can only send or submit things that are low-stakes and reversible.
The question is not "can the model do this?" It is "what happens when it does this wrong, and who notices?"
How to start
Pick one workflow your team already runs by hand every week. Write down each step, the systems it touches and the decision points. That document is your agent's first specification — and it often shows that half of the process can be automated with no AI at all.
Then build the smallest version, keep a human approving every action for the first few weeks, and only widen its permissions once you have seen it be right, repeatedly, on your real data.
Frequently Asked Questions
Is agentic AI just a chatbot with more features?
No. A chatbot responds to one message at a time. An agent pursues a goal across many steps, uses tools such as your LMS or calendar, and decides its own next action. The model inside can be the same; the loop around it is what makes it an agent.
Do we need a new platform to use AI agents in a course?
Usually not. Most useful agents connect to systems you already run — the LMS, a shared drive, email or chat. The work is in defining the goal, the allowed actions and the approval points.
Is it safe to let an agent message learners directly?
Start with the agent drafting and a person sending. Once you have reviewed enough of its drafts to trust it on a specific, low-stakes message type such as reminders, let it send those on its own.
