A data-gathering agent kicks off on the ADT event, pulls the record, and drafts a synopsis. Then each module is a specialist agent that files one standard report, and a coordinator agent synthesizes them into a single prioritized worklist for the nurse or NP.
The data agent runs first on the ADT event and hands the nurse a synopsis; the specialists each report; the coordinator synthesizes; the human reviews and can deep-dive into any agent.
The agents, and who does what
What kicks things off, what each agent does, how often they run, and what stays the human's job.
Data-gathering & monitoring agent
Runs first, then keeps watching
Kicks off before anyone opens the chart. On an ADT admit/discharge (or when the nurse opens the case) it:
ingests whatever records are available — discharge summary, labs, meds, prior notes, claims
flags the triggers it finds in the record, as suggestions
drafts a one-screen synopsis for the nurse to review
keeps monitoring for new ADT events and new data, and re-triggers the pipeline
Each module agent
Reviews one problem area
A specialist for its disease or risk (fluid, arrhythmia, diabetes, falls…). On its slice of the record it:
intakes the recent hospital events
reviews what's already known
tests whether it has enough data and enough context to reason
files findings, next-best-actions with a confidence, urgent escalations, and cross-module signals
Cadence & triggers
Wakes on events, not constant polling
Agents sit idle until something changes. The pipeline re-runs when:
an ADT event posts (admit or discharge)
new records land (labs, claims, care-gap feed)
the nurse opens or refreshes the case
a data gap gets filled — re-waking that agent and the coordinator
Post-discharge: fires on the ADT event. CHE: fires across all records when the annual review opens.
Coordinator agent
Keeps the running picture
Listens to every agent's report and re-synthesizes whenever one changes. It:
dedupes the shared threads (one syncope, not four problems)
resolves conflicts (anticoagulate the AF vs. hold for the bleed → surface, don't decide)
ranks by urgency × certainty into one worklist
labels each item: cleared / needs data / needs context / actionable
The human
Reviews and decides
The nurse or NP starts at the synopsis and the coordinator's output, not the raw modules. They:
act on urgent escalations first
fill the requested data/context — certainty rises live as they do
deep-dive into any agent to see its reasoning and evidence
accept or override every suggestion
The system routes and gathers; the human decides.
Walk through a real case
Demo case: an in-center HD patient is discharged after a syncopal episode. The ADT event fires the data agent; a VBC nurse then works the case. Watch the pipeline fill in.
Step 1 of 12
Live pipeline state
Data agent — idle
Record synopsis — auto-drafted
Detected triggers:
Syncope / fallHypoglycemia
Key facts: BP 98/56 post-HD · HR 48 irregular · K 3.4 · Hgb 9.1 · AF · on beta-blocker · in-center HD
Falls
asleep
Arrhythmia
asleep
Fluid
asleep
Diabetes
asleep
Coordinator — idle
Prioritized worklist
— appears once the coordinator has run —
What's a mock today, what's real in production
The report contract is the durable part. The reasoning engine behind each agent swaps without changing anything downstream.
This mock. The data agent's synopsis and each specialist are deterministic rules over the module's declared data/context needs. Great for showing the flow and the contract.
Production. The data agent ingests live ADT/records; each specialist becomes an LLM agent with tool access to the chart, labs, and claims — emitting the same report contract.
Unchanged either way. The coordinator, the worklist, the human review, and the deep-dive all read the contract, not the engine. That's the v2 → v3 path.
Companion explainer for the Complex Multi-Disease Framework · designed to serve as the tool's opening screen.