The Problem

Every year, 2.3 million Americans are harmed during care transitions — the moment a patient moves from hospital to home. Not because anyone failed, but because safe discharge requires 6 complex workflows happening simultaneously:

  • Clinical handoff to the receiving provider
  • Medication reconciliation and safety review
  • Insurance prior authorization for medications
  • Patient education in their native language
  • Social determinants of health screening
  • Post-discharge follow-up coordination

Today, every single one of these is done manually, serially, by different humans who rarely communicate. The result: a 20% readmission rate costing $26 billion annually.

This is not a data problem. All the data exists in FHIR. This is a coordination and synthesis problem — exactly what multi-agent AI was built to solve.

The Solution: ATLAS

ATLAS (Autonomous Transition & Longitudinal Agent System) is a 6-agent orchestration system built on MCP + A2A + FHIR that collapses a 6-hour manual process into 90 seconds.

Architecture

When a discharge is triggered, the ATLAS Orchestrator reads the patient's FHIR record via Prompt Opinion's FHIR context extension and dispatches 6 specialized agents:

Agent 1 — Clinical Handoff (MCP) Reads FHIR patient data and generates a specialty-tailored handoff letter. A cardiology patient gets a different handoff than a primary care patient — vocabulary, depth, and priorities all shift based on the receiving provider's specialty.

Agent 2 — Medication Safety (MCP) Reconciles pre-admission and discharge medications. Flags high-risk drugs (warfarin, insulin, opioids), dangerous interactions, renal dose adjustments for CKD patients, and medications requiring prior authorization.

Agent 3 — Prior Authorization (MCP) Drafts complete prior auth letters with clinical justification — the kind of evidence-backed arguments that anticipate insurance denial criteria. This is impossible with rule-based systems.

Agent 4 — Patient Navigator (MCP) Generates personalized discharge instructions at 6th grade reading level in the patient's preferred language — automatically detected from the FHIR Patient resource. Our demo patient Maria Garcia receives her instructions in Spanish, automatically.

Agent 5 — SDOH Screener (MCP) Screens for social determinants of health — transportation, food security, housing, financial toxicity, caregiver support. Generates community resource recommendations and flags for the care team.

Agent 6 — Loop Closure (MCP) The differentiator. 48 hours after discharge, this agent checks: Did the patient fill their prescriptions? Was the follow-up appointment scheduled? Were there any ED returns? It generates a personalized SMS outreach in the patient's language and alerts the care manager.

Why AI — Not Rules

Every output ATLAS generates requires multi-step reasoning and natural language generation that is fundamentally impossible with rule-based software:

  • Tailoring handoff depth to a specialist's needs requires understanding medical context
  • Writing prior auth letters that anticipate denial criteria requires reasoning about insurance policy and clinical evidence
  • Generating culturally appropriate instructions in the patient's native language requires generative AI

Technical Implementation

  • Platform: Prompt Opinion (MCP + A2A + FHIR)
  • Protocol: MCP JSON-RPC 2024-11-05 with FHIR Context Extension
  • FHIR: Full SMART scopes — Patient, Condition, MedicationRequest, Observation, AllergyIntolerance, MedicationDispense, Appointment
  • AI: Groq API (Llama 3.3-70b) for all clinical generation
  • Deployment: 6 FastAPI servers on Render.com
  • A2A: Orchestrator delegates to ATLAS Discharge Agent via A2A
  • Marketplace: 6 MCP servers + 1 A2A agent published

Impact

  • 36 million hospitalizations per year in the US
  • Every discharge could use ATLAS
  • 20% readmission rate → ATLAS addresses primary drivers
  • $26 billion annual cost of preventable readmissions
  • 6 hours of clinician time → 90 seconds with ATLAS

Feasibility & Safety

  • All FHIR access through authorized patient sessions
  • Outputs labeled as AI-assisted drafts for clinician review
  • ATLAS augments clinicians — it does not replace them
  • FHIR R4 is federally mandated — works with Epic, Cerner, Meditech

Live Demo

Watch ATLAS process a complete discharge for Maria Garcia — a 68-year-old Spanish-speaking heart failure patient — in under 90 seconds, with all 6 agents coordinating in real time.

Built With

  • a2a-(agent-to-agent)
  • fastapi
  • fhir-r4
  • github
  • groq-api-(llama-3.3-70b)
  • json-rpc
  • mcp-(model-context-protocol)
  • prompt-opinion-platform
  • python
  • render.com
  • smart-on-fhir
Share this project:

Updates