MediCore AI - Multi Agent Healthcare Orchestrator

8 specialist AI agents covering the complete patient journey - intake, diagnosis, treatment, insurance, follow-up & FHIR memory - orchestrated in one unified clinical response via PromptOpinion A2A.

Comment

Inspiration

Healthcare AI today is fragmented.

Most solutions act as isolated chatbots — one tool summarizes notes, another suggests diagnoses, another handles scheduling — but real healthcare requires coordinated decision-making across multiple departments, systems, and specialists.

Clinicians constantly switch between EHRs, insurance portals, follow-up systems, and care coordination tools just to manage a single patient journey.

We asked a simple question:

What if one clinical workflow request could activate an entire healthcare coordination system automatically?

That idea became MediCore AI. We were also inspired by PromptOpinion’s A2A protocol and the possibility of true multi-agent orchestration — where specialized AI agents collaborate in real time instead of a single LLM pretending to handle every responsibility.

MediCore AI was built around one core belief:

Healthcare does not fail because of lack of intelligence. It fails because intelligence is disconnected.

What it does

MediCore AI is a distributed multi-agent healthcare orchestration platform that coordinates the full patient journey using interoperable AI specialists. A clinician or care coordinator submits a single clinical workflow request. Behind the scenes, multiple healthcare agents activate collaboratively through A2A orchestration and generate one unified coordinated clinical response.

The platform currently includes: Agent | Responsibility

MediCore Orchestrator | Central coordination and response synthesis Intake Agent | Symptom extraction and risk stratification Diagnosis Agent | Differential diagnosis and urgency analysis Treatment Planner Agent | Treatment plans and monitoring workflows Insurance Billing Agent | Cost analysis and affordability support Follow-Up Adherence Agent | Longitudinal adherence and intervention planning Care Navigator Agent | Referral coordination and continuity-of-care Health Memory Agent | Longitudinal FHIR patient context retrieval

Enterprise clinical decision coordination. Orchestrates 8 specialist agents in sequence:

Step Agent Purpose
1 health_memory_agent Longitudinal FHIR history
2 diagnosis_agent Clinical analysis
3 intake_agent ICD-10-CM coding
4 care_navigator_agent Care pathways + referrals
5 social_barrier_agent SDOH screening
6 treatment_planner_agent Evidence-based treatment
7 insurance_billing_agent Cost, coverage, prior auth
8 followup_adherence_agent Follow-up scheduling

MediCore AI integrates with Medplum using FHIR R4 standards, enabling all specialist agents to reason over structured patient records instead of temporary conversational memory. The final output combines:

  • clinical intake analysis
  • diagnosis reasoning
  • treatment planning
  • medication affordability analysis
  • adherence monitoring
  • coordinated care navigation
  • longitudinal patient context

into one consolidated healthcare workflow response. Architecture

Prompt Opinion
     │  POST /         → MediCore AI Orchestrator
     │  POST /memory   → Health Memory Agent
     │
     ▼
┌─────────────────────────────────────────────────────┐
│  ngrok tunnel  (single tunnel, two agents)          │
│  https://<ngrok>.ngrok-free.dev  → localhost:8003   │
└────────────────────┬────────────────────────────────┘
                     │
          ┌──────────▼──────────┐
          │  orchestrator/      │  POST /
          │  Express app        │  GET  /.well-known/agent-card.json
          │                     │
          │  app.use('/memory', healthMemoryApp)
          └──────────┬──────────┘
                     │
          ┌──────────▼──────────┐
          │  health_memory_     │  POST /memory
          │  agent sub-app      │  GET  /memory/.well-known/agent-card.json
          └──────────┬──────────┘
                     │
     ┌───────────────┼───────────────────────┐
     ▼               ▼                       ▼
OpenRouter      Medplum FHIR R4         specialist agents
(LLM calls)   (longitudinal memory)  (called in-process)

How we built it

We designed MediCore AI using a distributed orchestration architecture powered by PromptOpinion’s A2A external agent system.

Core Stack

  • PromptOpinion A2A orchestration
  • Node.js + Express
  • TypeScript
  • OpenRouter
  • Medplum
  • FHIR R4
  • ngrok
  • Docker Compose

Each healthcare specialist agent runs independently and exposes its own A2A-compatible endpoint. The MediCore AI Orchestrator acts as the central coordination layer responsible for:

  • request routing
  • specialist delegation
  • structured response synthesis
  • fallback handling
  • workflow consolidation

FHIR Integration

We integrated Medplum as the longitudinal healthcare memory layer. Instead of relying on short-term conversational memory, agents retrieve:

  • patient history
  • ServiceRequests
  • medication context
  • longitudinal clinical records directly from structured FHIR resources.

Architecture Optimization

One major architectural decision was reducing the entire system to a single public ngrok tunnel. Rather than exposing multiple external tunnels, we mounted the HealthMemoryAgent as an Express sub-application inside the Orchestrator instance.

This allowed:

  • one HTTPS endpoint
  • independent A2A agents
  • simpler orchestration
  • cleaner deployment
  • more production-ready infrastructure

A2A Workflow

The orchestration pipeline works as follows:

  1. PromptOpinion sends a clinical workflow request
  2. MediCore AI Orchestrator receives the request
  3. Relevant specialist agents activate through A2A communication
  4. Agents retrieve longitudinal FHIR patient context
  5. Structured outputs are returned to the Orchestrator

6. The Orchestrator synthesizes one coordinated healthcare response

Challenges we ran into

One of the biggest challenges was stabilizing A2A communication between PromptOpinion and distributed external specialist agents.

We encountered issues involving:

  • endpoint synchronization
  • ngrok routing
  • schema mismatches
  • orchestration failures
  • agent registration reliability
  • structured response consistency

Another major challenge was designing healthcare workflows that felt operationally realistic rather than generic AI outputs. We wanted the system to reason not only about diagnosis, but also:

  • affordability barriers
  • transportation issues
  • missed follow-ups
  • medication non-adherence
  • continuity-of-care breakdowns

FHIR integration was another major engineering challenge.

Propagating patient context cleanly across multiple specialist agents without data leakage required designing a shared healthcare context layer across the orchestration pipeline.

Accomplishments that we're proud of

We are proud that MediCore AI evolved beyond a traditional healthcare chatbot into a real distributed clinical orchestration platform.

Key accomplishments include:

  • Fully functional multi-agent A2A orchestration
  • Real FHIR-native interoperability through Medplum
  • Longitudinal patient context retrieval
  • Distributed specialist AI collaboration
  • Coordinated continuity-of-care workflows
  • Production-style orchestration architecture
  • Structured multi-specialist healthcare outputs

We are especially proud that MediCore AI addresses operational healthcare problems often ignored by AI demos, including:

  • affordability analysis
  • adherence monitoring
  • insurance barriers
  • follow-up coordination
  • transportation support
  • care navigation workflows

The platform demonstrates how AI agents can augment real healthcare coordination instead of simply generating isolated text responses.

What we learned

This project taught us that multi-agent systems become dramatically more powerful when combined with interoperable healthcare infrastructure. We learned:

  • how to orchestrate distributed AI services
  • how real A2A communication works in production-like environments
  • how FHIR enables longitudinal healthcare reasoning
  • how healthcare workflows extend far beyond diagnosis generation

One of our biggest insights was that:

System prompt engineering is architecture. The quality of orchestration depends heavily on:

  • specialist role separation
  • structured outputs
  • context propagation
  • fallback handling
  • orchestration reliability We also learned that constraints often improve architecture.

The single-tunnel limitation pushed us toward a cleaner, more scalable orchestration design than we initially planned.

What's next for MediCore AI — Multi Agent Healthcare Orchestrator

Our next goal is transforming MediCore AI into a scalable clinical orchestration ecosystem.

Planned future features include:

  • SMART on FHIR authentication
  • Epic and Cerner integration
  • voice-based clinician intake
  • provider dashboards
  • predictive risk scoring
  • real-time adherence alerts
  • autonomous care coordination workflows
  • multi-patient population health monitoring
  • secure healthcare deployment infrastructure

We also plan to expand specialist coverage into:

  • cardiology
  • oncology
  • mental health
  • chronic disease management
  • hospital discharge coordination

Ultimately, we envision MediCore AI as:

A distributed clinical intelligence layer for coordinated healthcare delivery at scale.

Built With

Share this project:

Updates