What is an agentic LMS, and the one question that separates them
Agentic learning platforms let AI agents do multi-step work by themselves. The difference that decides procurement is whether the agent applies that work, or hands it to a person.
Every learning platform will describe itself as agentic within the year, so the word is already close to useless on its own. One question restores the signal, and it is not about how capable the agents are. It is: when an agent finishes a piece of work, does it apply that work itself, or does it hand it to a person? Everything an institution has to sign off on follows from the answer.
Agentic LMS
A learning platform in which AI agents carry out multi-step work on their own — drafting a course from source material, marking a submission against a rubric, assembling a report — rather than answering one prompt at a time. The term describes how much of a task the system will take on unprompted. It says nothing, by itself, about who is accountable for the result.
The category arrived quickly and from the top of the market. Docebo announced AgentHub at its Inspire event on 21 April 2026, describing agents that "reason, decide, and act" across enterprise knowledge and skills, with full availability planned for the autumn. Absorb published a definition of an agentic learning system on 5 June 2026 — agents that "take action across learning workflows" — and reported that in its Aura beta the architecture resolved 40 to 60% of routine admin support tickets "with no human opening a queue". A run of 2026 buyer's guides now rank agentic platforms as a category of their own.
What is the difference between an AI-powered LMS and an agentic one?
AI-powered describes features bolted to an existing product: a summarise button, a question generator, a chat panel. Each does one thing when asked. Agentic describes scope — the system takes on a whole task across several steps, decides what to do next inside that task, and keeps going without being prompted at each turn. The move from one to the other is real. It is also the point at which the question of who applies the result stops being academic.
The split inside the category
That Absorb figure is worth sitting with, because it argues against a lazy version of this post. Deflecting routine enrolment and certification questions is exactly where autonomy earns its keep: the questions are repetitive, the answers come from approved content, and a wrong one costs somebody a second email. Nobody needs a person in front of that. The stakes are not evenly distributed across a learning platform, and a policy of "a human approves everything" applied uniformly is just a slower product. The argument is narrower than that — it is about the handful of surfaces where being wrong is expensive.
Two designs are being sold under the same word, and they behave very differently the first time something goes wrong.
- Agents that act. The agent completes the task and applies the result. A course is published, an enrolment is assigned, a report is filed. You find out afterwards, usually from a notification. This is what most of the category means by agentic, and it is what the demos are built to show.
- Agents that propose. The agent completes the same task and stops. What it produced arrives as a draft, a queued item, a suggested score — and a named person applies it or does not. The work is identical; the last step belongs to somebody who can be asked why.
The second is slower. That is a real cost and worth stating plainly: if what you want is an assistant that simply does the thing, an approval step will irritate you. What you buy for the delay is that a model error is a rejected draft rather than an incident — a distinction that only matters occasionally, and matters enormously when it does.
The agent produces the work and hands it over. A person applies it, and the record keeps what was proposed beside what was accepted — including what was rejected, which is the entry proving the control was doing something.
The agent is instructed to check in before important actions, with importance judged by the model, in the same context window as the material it was asked to read.
That second pattern is worth naming because it is common and it sounds like the first. A system where the model is told to ask permission is not the same as a system where the model has no way to proceed without it. The first is a policy the model may follow. The second is a wall. Ask which one you are being shown, and ask to see the wall.
Why this is a procurement question and not a preference
For a company automating internal compliance reminders, autonomy is mostly an efficiency decision. For a school, a university or a regulated employer, it is a governance one, and increasingly a legal one. Legislatures have started writing rules about where AI may sit in assessment — several US states have moved on whether AI can be the primary basis for a grade, and the position is still changing. FutureEd's legislative tracker is the practical place to check what applies to you this term rather than last.
Even where nothing is written down yet, the question arrives at the same point in every institutional purchase. Somebody signs a paper saying what the system may do unattended. That person needs a shorter answer than a description of the model's instructions.
Can AI grade student work?
It depends on what you mean by grade. Producing a suggested score against a rubric, with the reasoning written out, is well within reach and saves real time. Releasing that score to a student as their result is a different act with different consequences, and it is the one rules are being written about. Ask a vendor to separate the two on the call. If the answer treats them as one thing, that is the answer.
Do agentic AI systems replace teachers and instructors?
Not in either design, though the two fail differently on the question. An agent that acts still needs somebody accountable for what it did — the work moves, the responsibility does not, and the person finds out later. An agent that proposes keeps that person in front of the decision instead of behind it. What actually goes away in both cases is the part of the job nobody defends: producing a first draft, marking the fortieth submission against the same rubric, rebuilding the same report every month. The judgement about whether a draft is right, a grade is fair, or a number means what it appears to mean is the part being protected, and it is the part institutions are paying for.
Six questions to ask an agentic vendor
These are ordered by how much they tend to reveal, and the first one is worth the whole call.
- List every action an agent can take without a person. Ask for it in writing, not on a slide. The length of the pause before the answer is data.
- Is confirmation a prompt or a check? If the model decides when to ask permission, permission is advisory. If the agent has no path to the live object, it is a control.
- Who releases a grade? Not who can override one afterwards — who releases it. These are different systems.
- What does the record show? A log of what the agent did is weaker than a record holding what was proposed, what was applied, by whom, and what was rejected.
- Can we turn the agents off? Per organisation, not per tenant. A group often needs one school running with them and another without.
- What happens on a bad day? Ask what a wrong answer costs and who finds out. "You correct it afterwards" and "it never reached anyone" are different products.
Where Lurno sits, and what it costs us
Lurno is the second column. Its agents draft lessons and questions from documents you upload with the source shown beside each point, suggest grades against your rubric over an anonymised submission, tutor learners through escalating modes, and turn a plain-language question into a report. None of them can apply its own work, and that is not a setting — there is no path from an agent to a released grade or a published course. The most consequential actions need two different people, and neither may be the one who proposed the change.
The reporting agent is the clearest case. It writes the report definition and never reads your data. What it proposes runs back through the same permission checks as a report somebody built by hand, so it cannot invent a number and cannot show an organisation the person asking was not already allowed to see.
- The delay is real. There is no mode in which our agents write to a live programme or release a grade by themselves. If you want autonomy, we are the wrong shortlist entry and it is better to know now.
- SCORM, xAPI and LTI are in development. Modelled in the product, runtime still being built. Courseware locked in SCORM packages is not something we can ground an agent in today.
- Agents can be switched off entirely, per organisation. Everything else keeps working — they are additive, not load-bearing.
Is this the same as human-in-the-loop AI?
Nearly, though the phrase undersells it. Human-in-the-loop usually describes a review step added on top of a system that could otherwise run alone — the autonomy exists and a checkpoint is placed in front of it. Here the approval is the architecture: nothing an agent produces reaches a learner, a grade or a report without someone accepting it, so there is no autonomous path underneath to review. The practical difference shows up when the review step is under load. A checkpoint can be skipped. A missing path cannot.
Does an approval step defeat the point of agentic AI?
Only if the point was removing people. The expensive part of the work is not the click that accepts a draft — it is producing the draft, marking against the rubric, assembling the report. An agent that does that and stops has taken the hours and left the judgement. For most institutions that is the trade they wanted, and several are only permitted to make that one.
If you are evaluating agentic platforms, take question one to all of them and compare the written answers. We will give you ours before the call rather than during it, and if it rules us out, that is a useful hour saved. Book a demo, or read what each of our agents may do on its own on the AI page.