About the Project

Healthcare interoperability solved data exchange.

It did not solve intelligent coordination between systems, workflows, patients, providers, and now AI agents.

That is the problem AgentInterop was built to explore.

AgentInterop is an open interoperability framework for healthcare AI agents built on top of:

  • HL7 FHIR
  • MCP (Model Context Protocol)
  • SMART-on-FHIR concepts
  • Agent orchestration patterns
  • Human approval workflows
  • Provenance-aware healthcare interactions

The project explores how healthcare agents can securely discover each other, exchange context, coordinate actions, and collaborate across fragmented healthcare ecosystems using open standards instead of proprietary silos.

The live prototype UI demonstrates a healthcare-native multi-agent environment designed around interoperability first. (Agents UI)

GitHub Repository: AgentInterOp GitHub Repository

Live Demo: AgentInterop UI Demo


The Why

AI agents are rapidly becoming capable of:

  • reasoning over clinical information
  • coordinating workflows
  • summarizing records
  • assisting with care management
  • automating administrative operations

But healthcare systems remain fragmented.

FHIR standardized APIs, but healthcare still lacks a shared interoperability layer for autonomous systems.

Today:

  • agents cannot safely collaborate across organizations
  • context is isolated
  • workflow orchestration is brittle
  • trust and governance are inconsistent
  • healthcare AI tooling is mostly disconnected point solutions

AgentInterop explores what happens when healthcare agents become interoperable by design.

The goal is not replacing clinicians.

The goal is reducing fragmentation and enabling intelligent coordination.


What AgentInterop Does

AgentInterop provides a framework where healthcare agents can:

  • Discover capabilities dynamically
  • Exchange structured requests
  • Coordinate workflows
  • Share healthcare context
  • Route tasks intelligently
  • Maintain provenance trails
  • Enforce human approval checkpoints
  • Operate against FHIR-native resources

The project introduces the concept of healthcare agents acting more like interoperable participants inside a shared ecosystem instead of isolated AI applications.


Example Healthcare Agents

The platform explores specialized agents such as:

Clinical Reasoner

Analyzes patient context and proposes evidence-based next actions.

Care Coordinator

Orchestrates follow-ups, transitions of care, and workflow handoffs.

Patient Communication Agent

Supports patient engagement and longitudinal communication.

Data Retrieval Agent

Aggregates patient information across FHIR endpoints.

Privacy Guardian

Applies consent enforcement, audit controls, and policy evaluation.

Monitoring Agent

Tracks longitudinal health events and detects changes over time.

Rather than building one monolithic AI system, AgentInterop explores distributed intelligence where agents collaborate through standardized healthcare interactions.


Core Architecture

AgentInterop is built around several major concepts:

FHIR-Native Communication

Agents communicate using healthcare-native data structures instead of generic JSON payloads.

This includes:

  • Patient
  • Observation
  • Encounter
  • CarePlan
  • Condition
  • MedicationRequest
  • Task
  • Provenance
  • Communication

resources as shared context layers.


Agent Discovery + Routing

The system explores dynamic discovery of agent capabilities so workflows can route requests to the most appropriate healthcare agent.

Examples:

  • care gap analysis
  • summarization
  • patient outreach
  • utilization review
  • quality measurement support
  • longitudinal risk analysis

Human-in-the-Loop Safety

Healthcare AI cannot operate safely without oversight.

The platform incorporates:

  • approval checkpoints
  • audit logging
  • provenance tracking
  • explainable interactions
  • permission-aware orchestration
  • step-up authorization patterns

This concept was heavily influenced by the broader HealthClaw guardrail framework.


MCP + Interoperability

The project experiments heavily with MCP (Model Context Protocol) concepts for structured tool interoperability between healthcare agents and systems.

The broader AI ecosystem is rapidly moving toward open agent tooling standards and interoperable agent architectures. (Vercel)

AgentInterop applies those ideas specifically to healthcare.


Technical Stack

The project combines:

  • Next.js
  • React
  • TypeScript
  • Tailwind CSS
  • Vercel deployment workflows
  • MCP patterns
  • FHIR APIs
  • Agent orchestration concepts
  • Healthcare authorization models
  • Provenance-aware workflows

The architecture is intentionally modular and extensible to allow experimentation with evolving healthcare AI standards.


Inspiration

This project came from years of working in:

  • healthcare interoperability
  • FHIR analytics
  • payer/provider exchange
  • quality measurement
  • healthcare data quality
  • AI-assisted healthcare tooling

A recurring pattern became impossible to ignore:

Healthcare organizations still spend enormous effort moving information between disconnected systems while patients and clinicians deal with fragmented experiences.

At the same time, AI systems are becoming increasingly capable of coordinating complex workflows.

The missing layer is interoperability between intelligent systems.

FHIR standardized healthcare data.

AgentInterop explores standardizing healthcare agent collaboration.


Challenges

Some of the hardest challenges included:

  • balancing autonomy with clinical safety
  • handling inconsistent FHIR implementations
  • creating explainable agent behavior
  • designing interoperable workflows
  • preserving healthcare privacy constraints
  • modeling trust between agents
  • maintaining human oversight
  • preventing unsafe autonomous actions

Healthcare is one of the most difficult environments for agent systems because:

  • trust matters
  • provenance matters
  • auditability matters
  • explainability matters
  • humans must remain in control

What I Learned

This project reinforced several major lessons:

  1. FHIR APIs alone are not enough.
  2. Healthcare needs interoperable intelligence layers.
  3. Agent ecosystems require governance and trust.
  4. Human oversight must remain central.
  5. Open healthcare ecosystems will likely outperform closed AI silos long term.

Most importantly:

Healthcare does not just need smarter AI.

It needs infrastructure that allows humans, systems, and intelligent agents to safely work together.

Built With

  • a2a-protocol
  • anthropic-claude
  • cql-(clinical-quality-language)
  • fastapi
  • fhir-r4
  • flask
  • gunicorn
  • jinja
  • json-rpc-2.0
  • model-context-protocol-(mcp)
  • pydantic
  • python
  • server-sent-events
  • smart-scheduling-links
  • vercel
Share this project:

Updates