The Care Plan That Writes Itself, In A Demo

A preprint from four University of Missouri researchers describes AgenticAD, a system of eight AI agents in which three of them interview a caregiver and produce an assessment summary, a daily plan and a follow-up schedule. It has been run on three invented personas and no real families.

Every family that lives with Alzheimer's eventually assembles the same private dossier: the night he got up at 3am and could not find the bathroom; the week she stopped finishing meals; the fall in the hallway that nobody quite saw happen. It exists in text messages, on the backs of envelopes, in a caregiver's head. Then comes the fifteen minute appointment, and somebody has to turn all of it into a paragraph a clinician can act on. That translation problem, from lived chaos to structured summary, is the gap a preprint posted to arXiv on 1 September 2025 by Adib Bazgir, Amir Habibdoust, Xing Song and Yuwen Zhang tries to close with software.

Their framework, AgenticAD, is a blueprint for eight specialised, interoperable AI agents grouped into three functions: caregiver and patient support, data analysis and research, and advanced multimodal workflows. The one that matters for this section is the Alzheimer's Support Agent. It is not a single chatbot but a small committee: a swarm built on Microsoft's Autogen framework containing an Assessment Agent, a Care Plan Agent and a Follow-up Agent, all running on GPT-4o, behind a Streamlit interface. The user fills in structured fields (age, symptoms, living situation, safety concerns). The Assessment Agent turns those fields into a clinical summary. That summary is handed to the Care Plan Agent, which proposes practical daily plans, safety mitigations and clinical next steps. The Follow-up Agent closes the loop. The output arrives in three named parts: an Assessment Synthesis, a Safety-First Daily Care Plan, and a Follow-up and Monitoring Plan.

What did it actually produce? The authors report running the agent against three user personas: "Clinician / Staff", "Caregiver / Family Member" and "Person with Memory Concerns". In the caregiver scenario, a family member whose stated goal was establishing a routine for a 76-year-old with paranoia and aggression, the generated plan emphasised monthly medication reviews for donepezil and tracking of agitation episodes. In the clinician scenario, focused on safety for a 76-year-old living alone with poor judgment and frequent falls, the output was framed more clinically, highlighting medication adherence for memantine and offering sensitive prompts for raising the subject of driving. These are descriptions of what a language model wrote in a demonstration. They are not recommendations, and no reader should read a drug name in a research paper as a suggestion about anyone's own tablets.

The most interesting piece, for families, is the least glamorous. The Follow-up Plan came out consistently structured across all the test cases, organised under five headings. The check-in cadence recommended weekly notes and monthly medication and behaviour reviews. The tracking template listed the metrics to watch: sleep hours, agitation episodes, falls, meals, medication adherence. The escalation criteria set out when to phone a clinician and when to go to urgent care or an emergency department, keyed to signs such as acute confusion and new neurological deficits. If you have ever tried to work out at midnight whether tonight is a phone call or an ambulance, you understand why a legible, consistent version of that page is worth something.

Now the boundary. This is a preprint, which means it has not been through peer review. It is explicitly a methodological framework, an architectural blueprint, and the evaluation is a walkthrough of what each of the eight agents produced when the authors exercised it. There were no real caregivers, no patients, no reported clinician ratings of whether the summaries were accurate or the plans safe, no comparison against usual care, no error rate, no measure of what happens when a distressed relative types the wrong thing into the intake box. The authors are candid about this. Their conclusion states that the next frontier is technical integration and clinical validation, and they note that agents relying on external tools showed operational dependencies and failure modes that argue for fault-tolerant engineering. The paper's own claim is viability of the architecture, not benefit to anyone.

So the honest reading is this: nobody has shown that AgenticAD helps a single family, and nobody has shown it is safe to put in front of one. What the paper does is sketch the shape of the tool that families keep improvising for themselves, and it puts that sketch somewhere other researchers can test it, break it and argue with it. That is a modest thing. It is also how the useful ones start.