Topics discussed:

• Sam walked a fictional 76-year-old with a hip fracture, diabetes and obesity from hospital into home care. Agents matched her against the FHIR server, pulled her existing conditions and vitals through tools, filled out an admission profile conversationally, then handed it to a care plan writer that consults specialist expert agents — diabetes, obesity, home care, and notably not hypertension, because she doesn't have it. The output is a FHIR CarePlan, with each activity also rendered as a BPMN diagram and deployed to an automation engine that then drives the nurse's home visit deterministically. The reason for the diagrams is a hypothesis worth testing: a clinician can spot and correct a problem in a picture faster than in a block of prose describing what the system will do. Asked to raise a CGM alert threshold to above 200 for more than two hours, the system re-consulted the other experts for cross-dependencies before changing anything; and an AI watching the home visit later noticed a high blood pressure reading and proposed adding daily BP monitoring as a suggestion rather than an edit — it doesn't move the plan out from under anyone.
• Sam's follow-up question was the honest one: those expert agents were prompts he hand-wrote, and he isn't a clinician. So can they be generated from clinical practice guidelines instead? The workflow extracts the decisions from a guideline as DMN — a reviewable decision table any rules engine can execute — and builds a care plan writer per guideline that calls those decisions over MCP. The hypertension guideline wants more than two qualifying readings before treatment; given a note with a single 140/90, the agent correctly recommended obtaining out-of-office measurements rather than starting treatment. Provenance resources mark what was AI-asserted and link each activity back to its source guideline.
• Malte Sussdorff raised the objection the demo invites: agents visibly making clinical suggestions and writing a plan that a doctor merely reviews looks a lot like a clinical decision support system, which drags you into medical device registration. Sam agreed it's fair and argued the mitigation is to push as much as possible into deterministic DMN and BPMN, shrinking the part where a non-deterministic model is making a judgement — and to pick use cases like discharge planning or home care admission, where no diagnostic decision is being made at all. Malte described the line his own team draws: an LLM can show a doctor possible diagnoses, but nothing reaches the FHIR store until the doctor reads it, decides, and re-enters it — the doctor has to be the decision-maker of record, not a rubber stamp.
• Sam's broader claim: using AI to generate more narrative is missing the point — you're adding another document for another AI to summarise later, when the value is in actionable output like BPMN processes and structured FHIR. He'd also like FHIR to represent things like a differential diagnosis, since a note exists partly to capture the doctor's thinking, and a model can only capture what it thinks that thinking was. Giulia Baldini pushed back from the European side: clinicians still need narrative — the discharge letter has to look a certain way — and the people she's worked with won't adopt this quickly, so something is needed in the middle. Sam agreed on the timeline while wanting to skip the narrative layer wherever possible.
• Olim demoed Atomic Workspace, where the analysis layer is a directed acyclic graph of SQL on FHIR view definitions — flatten patient properties in one view, drill further in the next — plus a FHIR-to-OMOP translation underneath. Asked to compare diabetic and hypertensive patients, the agent went to the OHDSI phenotype library, imported two cohort definitions, converted them from CIRCE into SQL and charted the result in Vega-Lite; cohorts are stored with provenance tags and their original representation so you can see where each came from and rebuild it, and the agent gets tools to look up standard vocabulary codes rather than hallucinate them. Nikolai noted the FHIR-to-OMOP translation is open source under the SQL on FHIR subgroup and passes the published golden tests. Why bother with OMOP at all: as Arjun put it, it's the dominant standard in clinical research — 20-plus years, tens of thousands of papers, diagnoses in SNOMED and medications in RxNorm — so a cohort definition from any paper runs against any OMOP dataset, and SQL on FHIR defining those views brings that whole ecosystem within reach of agentic tools.