Darena Health Agents

Inspiration

Hospital readmissions cost Medicare $26 billion every year, and studies show that up to 30% of readmissions in elderly patients trace back to a single preventable failure at the moment of discharge: nobody performed a rigorous medication reconciliation. Polypharmacy kills an estimated 128,000 Americans annually — more than stroke.

For example: A 72-year-old walks out the door holding eleven prescriptions, three of which interact, two of which raise her bleeding risk, and one of which her pharmacy will give her for the next five years without a single clinician looking at the combined list again.

This is not a data problem. The data is all in the chart. This is a reasoning problem — and it's exactly the kind of problem that belongs to a new generation of autonomous AI agents.

What it does

Darena Health Agents is a three-agent AI system that autonomously performs medication safety review and discharge readiness assessment for hospitalized elderly patients — built natively on the three protocols at the heart of the hackathon: MCP, A2A, and FHIR.

Meet the agents

🔴 RxSafe — Medication Safety Reviewer Reads the patient's full FHIR medication list, identifies polypharmacy risks, quantifies anticholinergic burden, flags dangerous drug–drug interactions, and cross-references against Beers Criteria. Exposed as a public A2A endpoint so any other agent can call it.

🟠 DischargePlanner — Care Transition Orchestrator When discharge is considered, DischargePlanner autonomously invokes RxSafe over the A2A protocol — no human in the loop, no button click — synthesizes the medication safety findings into a structured discharge care plan, and surfaces anything that requires physician sign-off before the patient leaves the hospital.

🟢 DischargeReady — The Judgment Engine Analyzes four days of time-series vital signs, lab trends, nursing observations, and physical therapy assessments to decide whether the patient is READY / CONDITIONAL / NOT READY for discharge — with a confidence level, specific evidence citations from the chart, and explicit identification of ambiguities that still require human physician judgment.

Where is the AI?

This is the question every hackathon judge asks, and we built the project specifically to answer it. We tested the system on Margaret Chen, a synthetic but clinically realistic 72-year-old with 11 active medications, 55 time-series observations, and 4 days of inpatient data.

DischargeReady did four things a rules engine fundamentally cannot:

1. Predicted orthostatic hypotension risk from the combined pattern of blood pressure trends and diuretic therapy — a signal not visible in any single measurement.
2. Detected self-medication behavior by inferring it from the presence of OTC agents on the home medication list.
3. Flagged a triple bleeding risk that none of the individual drugs would have triggered on their own.
4. Refused to clear the patient for discharge — and explained, in full clinical prose, exactly why, and under what conditions she could safely go home.

This is not FHIR-as-a-database with an LLM stapled on. It is genuine cross-domain clinical reasoning over real longitudinal patient data.

How we built it

• Platform: Prompt Opinion's BYO Agents infrastructure
• Protocols: MCP (tool access), A2A (inter-agent autonomous collaboration), FHIR R4 (clinical data interoperability)
• Data: Custom FHIR R4 Bundle with MedicationStatement, Observation, DiagnosticReport, Condition, and AllergyIntolerance resources for a realistic polypharmacy case
• Agent design: Each agent is a specialized clinical role — medication pharmacist, discharge coordinator, attending physician — with its own system prompt, tool access, and output schema

Every Prompt Opinion BYO Agent is auto-exposed as a public MCP URL and A2A URL, so the agents can be called by any external system — EHRs, other agents, clinician-facing UIs — without us writing any protocol plumbing.

Challenges we ran into

• FHIR ingestion quirks: The platform initially returned empty medication data despite expanded OAuth scopes. We resolved this by restructuring the bundle into a proper polypharmacy case with current timestamps and verifying import through direct FHIR REST API queries.
• A2A orchestration: Writing a system prompt that makes DischargePlanner autonomously delegate to RxSafe — without human nudging — required several iterations. The breakthrough was a mandatory-orchestration clause in the system prompt plus deterministic output schema.
• Proving clinical reasoning: Distinguishing AI judgment from pattern-matched templates. We solved this by designing test cases where the correct answer requires cross-domain synthesis unavailable to any single rule.

Accomplishments we're proud of

• Three production-ready agents deployed and publicly addressable via MCP and A2A
• Autonomous A2A collaboration validated end-to-end (DischargePlanner → RxSafe)
• Clinically valid AI judgment demonstrated on a realistic polypharmacy case, with a NOT READY verdict that a board-certified internist would sign off on
• Physician-designed throughout — not a developer's guess at clinical workflow

What we learned

That the protocol plumbing is not the hard part anymore. MCP, A2A, and FHIR give you interoperability for free. The hard part — and the part that matters — is designing agents whose clinical reasoning you would actually trust near a patient.

What's next

• Expand the patient cohort with CHF, post-op, and oncology discharge cases
• Integrate with a real EHR sandbox (Epic / Cerner FHIR endpoints)
• Add a fourth agent: CareHandoff, which drafts the PCP follow-up letter from RxSafe + DischargeReady outputs
• Open-source the agent system prompts so other clinicians can audit, critique, and improve them

---

Submitted by Ken Liao and Christy Liao, MD (candidate). Built for Agents Assemble: The Healthcare AI Endgame, April 2026.

Built With

• a2a
• agent-to-agent
• claude
• fhir
• fhir-r4
• healthcare-ai
• llm
• mcp
• model-context-protocol
• polypharmacy
• prompt-opinion