Embodied agents driven by large language models can now explain a medical pamphlet in fluent, friendly language. In our earlier study, patients liked such an agent, and they learned no more from it than from the printed pamphlet. My argument is that this is a design problem, not a model-size problem. A language model does not carry a teaching plan: it does not know which idea to teach first, what the patient already understands, when to ask instead of tell, or how to check whether anything stuck.
The model fills in the words; the layer decides what the words are for.
The patterns, actions and concept-map format were first built and tested in my K-12 work outside campus (
AI PBL Agent) and then transferred to patient education. Three things make that transfer reasonable: most of the teaching patterns are anchored in research with adults, so K-12 was where they were compiled and tried, not the limit of where they hold; patient materials are already written for middle-school reading levels; and patient education adds a safety shell on top (source locking, red-flag information that may never be softened), which is part of the contribution.
A between-subjects study compares the layered agent with an otherwise identical agent driven by a single structured prompt. The primary outcome is delayed, unassisted comprehension: what a patient still understands later, without the agent in the room. The study is being prepared now.