AI agents that draft. People who decide.
Lurno's agents do the multi-step work — build a course from your documents, mark a submission against your rubric, turn a question into a report. What none of them can do is apply their own work. Every one stops and waits for a person.
What an agentic LMS is, and the question to ask about one
Lurno's AI agents do four jobs. They draft course content — lesson text, module outlines, mixed resources and assessment items — grounded in documents you upload, with the page each point came from shown beside it. They tutor learners through four escalating modes, with a check on our side that will not let one give the answer away. They draft grades over an anonymised submission that an instructor confirms or overrides. And they turn a plain-language question into a report. Every one of the four ends the same way: a person accepts the work, or it does not count.
- Agentic LMS
- A learning platform where AI agents carry out multi-step work on their own — drafting a course, marking a submission, assembling a report — rather than answering one prompt at a time. The question that separates them is what happens next: whether the agent applies its own work, or hands it to a person to approve.
Two kinds of agentic AI, and why the difference decides procurement
Most platforms marketing agentic AI are selling autonomy — agents that act without waiting for a person at each step. That is a real design, and for an institution it is the harder one to sign off.
| Capability | Agents that act | Agents that propose |
|---|---|---|
| Who applies the work | The agent | A person, every time |
| When you find out | After it has happened | Before it happens |
| Marking | The agent can release a grade | Only a person can release a grade |
| Publishing a course | The agent can publish | Two people approve |
| What the record shows | What the agent did | What was proposed, and who accepted it |
| When it gets something wrong | You correct it afterwards | It never reaches anyone |
Lurno is the second column. Not as a setting that can be switched off — an agent has no path to apply its own work.
Authoring, tutoring, assistance
Three surfaces, one rule: each drafts and hands over. Pick one.
Draft a lesson, then argue with it
Authors start from the outcomes they already wrote. A two-phase pipeline handles anything bigger than a paragraph: it plans the shape of the lesson first — an explainer, then three questions of named types — then fills each slot. Assessment items are planned as a set the same way.
- Adjust with AI on any block: make harder, simplify, add a worked example, fix wording. It returns a before-and-after diff, rendered as the block will look, and you keep one side.
- Image generation and audio narration run through the same draft-then-accept flow.
A tutor that will not give the answer away
A stuck learner climbs a ladder. Reformulate rephrases the question. Hint returns the next hint in sequence, never the last one. Reveal is the only mode permitted to state the answer, and only after an attempt. Explain covers why the answer is right and where the learner's went wrong.
- A check on our side compares each reply against the question's answer key for the mode requested, so reformulate and hint cannot leak the answer even if the model tries. Containment is enforced after generation, not asked for in a prompt.
- A learner who withdraws AI-processing consent gets no tutor: the server sends nothing to any provider. Every turn is metered and logged.
Ask about your data. Confirm the change.
The copilot answers questions in plain language. It emits a validated query against a registry of certified metrics and their allowed dimensions — never SQL, never a table or column name — and runs it read-only under the caller's own permissions. It cannot show anyone a number they could not open themselves.
- An AI-written query passes the same validator a hand-built report passes. If the catalogue lacks the metric a question needs, the copilot refuses and names the nearest one.
- Ask for a change and it proposes one with its risk level attached. A person confirms; high-risk actions ask twice; publishing or archiving a programme goes to a second approver.
From a PDF on somebody's laptop to an answer with a citation
- 01
Upload
PDF, DOCX or PPTX, into an organisation's library or onto a programme. Long files parse in the background.
- 02
Parse into a versioned source
The file becomes structured text, heading path and page numbers intact. Re-upload gives a new version, not a silent replacement.
- 03
Chunk and embed
Each chunk keeps the heading it sat under and the pages it spans, so a citation points somewhere a reader can turn to.
- 04
Retrieve with both methods
Vector and keyword search run over the same scoped versions and their rankings are fused, so an exact policy reference is found alongside a paraphrase.
- 05
Answer, and show the source
The reply names document, version and page. Attach documents at programme level and every generation inside it inherits them.
What happens to your text before it reaches a provider
A learner's message, an uploaded PDF, a question typed into the copilot: all content, never instruction. Six gates enforce it.
Injection classification
Untrusted input is classified before the real call runs. Text trying to redirect the model is flagged rather than executed.
A randomised fence
Untrusted text is wrapped in a per-request tag it cannot guess, and that tag is stripped from the content. Truncation re-closes the fence rather than orphaning it.
Redaction before the call
Names, emails and identifiers are replaced before anything leaves for a provider. Grading runs over an anonymised submission; the map back to the learner stays on the server.
Checked before it is used
Every response is checked for the shape we expect. A malformed reply gets one attempt to correct itself and then fails, rather than landing in a course.
Two-tier content screening
A fast heuristic blocks unambiguous cases before spending a call; a lighter model handles the rest. It is written not to flag anatomy, disease or the history of war.
A trail from proposal to change
Requests, accepted drafts and rejected drafts are separate entries in the audit log that cannot be quietly edited. What the model suggested sits next to what a person applied.
What each agent may do on its own
| Capability | The agent does | A person does |
|---|---|---|
| Draft a lesson block | Writes the draft | Accepts or rejects it before it enters the programme |
| Adjust an existing block | Proposes a rewrite as a before-and-after diff | Keeps the before or the after |
| Generate assessment items | Plans the mix, then drafts each item | Accepts item by item; a rejected item is never placed |
| Tutor a learner | Replies in the mode the learner asked for | Nothing — the answer-key guard runs server-side |
| Grade a submission | Suggests per-criterion rubric scores over an anonymised submission | Confirms and releases the grade. The AI has no path to release one |
| Answer a question about your data | Emits a validated query, read-only, under your permissions | Reads the result. No write happens |
| Change something in a programme | Proposes the change and states its risk level | Confirms it; high-risk actions ask a second time |
| Publish or archive a programme | Proposes it and queues it | Two people: the one who asked, and a second approver |
There is no mode in which Lurno's AI writes to a live programme or releases a grade without a person confirming it.
Questions people ask about AI agents
See it grounded in your own material.
Bring a syllabus, a policy PDF or a rubric. We will upload it, ground a generation in it, and show you the citation.