Exploring agents in education, carefully

Education attracts agent proposals faster than almost any other sector, and for understandable reasons: enormous volumes of repetitive administrative work, chronic shortage of individual attention, and material that is largely text.

It is also the sector where being wrong compounds in ways that are hard to see, because the person affected is frequently not in a position to notice.

Where the boundary belongs

Our position is that agents in education belong on the preparation side and not the assessment side.

Preparing material, retrieving relevant sources, drafting feedback for an educator to review, assembling administrative summaries: these remove hours from a week and the failure mode is a person correcting a draft.

Deciding a grade, judging whether a submission is original, or making a determination that affects a student's progression: these are decisions where the failure mode is a consequence for someone with limited recourse, and they need a person with accountability attached.

This is not a technical limitation. It is a view about where accountability should sit, and we would hold it even if the systems were better.

What we actually study

How agents change an educator's week, measured rather than estimated. Not which tasks are notionally automatable but which ones were handed over, what came back, and how much of it needed rework.

How retrieval behaves across languages in educational material, which is where we see the clearest evidence that aggregate quality figures hide the cases that matter most.

Where the corpus gaps are. A great deal of what makes teaching work is not written down anywhere, and an agent asked about it returns the closest adjacent thing with a citation attached.

The failure we worry about

Not a wrong answer that someone catches. A fluent, well-sourced answer built on material that is out of date or subtly inapplicable, delivered to someone without the context to question it.

An agent that marks its own uncertainty is less impressive and considerably safer here than one that always produces something confident.

What we would advise an institution

Start with administrative preparation, not anything touching assessment. Give it a named owner. Write evaluation criteria from real past cases. Keep the audit trail, because you will be asked.

And treat data residency as the first question rather than the last, particularly where student records are involved.

Next

Yaju Labs covers the research programme, and the Open Development Community is free to join. Catalyst Grants support projects building agent infrastructure, with applications opening in September 2027.