A healthcare-specific coordination and governance protocol for context-centric, policy-governed multi-agent clinical workflows.
Proposed · v0.1 · Not yet an adopted standard
Author: Eng. Mohamed Salah Ramadan — CEO & Lead Architect, Sky Host Software
Reference implementation: AI-HOS — AI-Native Hospital Operating System
MAIP (Medical Agent Interaction Protocol) defines a healthcare-specific coordination and governance layer for interactions between autonomous and semi-autonomous AI agents in hospital and clinical environments.
The protocol focuses on the semantics that become important when an AI-generated clinical intent must move across multiple specialized agents while preserving patient context, authorization, human oversight, evidence, task state, and provenance.
| Capability | MAIP v0.1 |
|---|---|
| Patient / Encounter Context | ✓ |
| Clinical Intent | ✓ |
| Authorization | ✓ |
| Human Approval | ✓ |
| Agent Capability Discovery | ✓ |
| Task Lifecycle | ✓ |
| Evidence | ✓ |
| Provenance & Audit | ✓ |
| Action Class × Autonomy Level | ✓ |
| FHIR interoperability | ✓ |
| A2A interoperability | ✓ |
| MCP interoperability | ✓ |
Important: These capabilities describe the protocol model defined by the v0.1 specification. They do not imply that the current AI-HOS reference implementation implements every capability in full.
Modern healthcare systems already have standards for clinical data exchange and, increasingly, standards for agent and tool interaction.
The problem appears when these pieces must participate in a single governed clinical workflow.
Consider a physician instruction:
"Patient has severe abdominal pain and vomiting. Order CBC, CRP, LFT, creatinine, and abdominal ultrasound. Start Pantoprazole and Ketorolac."
A production workflow may involve a Doctor Agent, Laboratory Agent, Radiology Agent, Pharmacy Agent, authorization policies, human approval, clinical evidence, and an audit trail.
The challenge is not simply moving a message from one component to another. The system must preserve the meaning and governance of the clinical action as it moves through the workflow.
MAIP addresses this problem by providing a structured interaction model for:
- shared patient and encounter context
- explicit clinical intent
- authorization and policy constraints
- agent-to-agent delegation
- human approval and oversight
- evidence references
- task lifecycle and execution state
- provenance and auditability
CLINICAL INTENT
│
▼
┌─────────────────────┐
│ MAIP │
│ Interaction & │
│ Governance Layer │
└──────────┬──────────┘
│
┌───────────────────────┼───────────────────────┐
▼ ▼ ▼
CONTEXT AUTHORIZATION EVIDENCE
│ │ │
└───────────────────────┼───────────────────────┘
▼
TASK LIFECYCLE
│
┌──────────┴──────────┐
▼ ▼
HUMAN APPROVAL AUTONOMOUS ACTION
│ │
└──────────┬──────────┘
▼
EXECUTION
│
▼
PROVENANCE / AUDIT
MAIP is intended to sit alongside, not replace, existing interoperability technologies.
MAIP
Clinical Coordination
& Governance
│
┌────────────┼────────────┐
▼ ▼ ▼
FHIR A2A MCP
Clinical Data Agents Tools
- FHIR provides standardized representations and exchange mechanisms for healthcare information.
- A2A provides a general framework for agent-to-agent interaction.
- MCP provides a standardized mechanism for agent interaction with tools and resources.
- MAIP focuses on healthcare-specific coordination and governance semantics across clinical multi-agent workflows.
MAIP is therefore designed as a complementary protocol layer, rather than a replacement for FHIR, A2A, MCP, or existing hospital interoperability infrastructure.
MAIP v0.1 organizes a governed agent interaction around several core concepts.
The interaction represents why an action is being requested, not only which technical operation should be executed.
Clinical actions are associated with the relevant patient and encounter context so that downstream agents can reason about the same clinical episode.
Actions are subject to authorization and policy constraints before execution.
Higher-risk actions can require explicit human approval rather than unrestricted autonomous execution.
A clinical request is treated as a lifecycle rather than a single fire-and-forget message.
REQUEST
│
▼
VALIDATE
│
▼
AUTHORIZE
│
├──────────────► HUMAN APPROVAL ──────────────┐
│ │
▼ ▼
EXECUTE ◄──────────────────────────────── APPROVED
│
▼
VERIFY / RECORD
│
▼
COMPLETE
│
▼
PROVENANCE / AUDIT
Clinical decisions and actions can carry references to supporting evidence and decision context.
The interaction is designed to support an auditable chain showing what was requested, authorized, executed, and recorded.
MAIP v0.1 introduces an action-class × autonomy-level model intended to make the relationship between an action and the permitted degree of autonomy explicit.
Conceptually:
AUTONOMY LEVEL
┌────────┬────────┬────────┐
│ Assist │ Review │ Execute│
└────────┴────────┴────────┘
▲
│
ACTION CLASS
│
┌───────────────┼────────────────┐
│ │ │
Informational Clinical High-risk
│ action action
│ │ │
└───────────────┴────────────────┘
The purpose is to avoid treating every AI action as equivalent.
A clinical information lookup, a diagnostic order, and a high-risk intervention should not necessarily have the same authorization or approval requirements.
The normative details belong in SPECIFICATION.md.
A multi-agent clinical environment needs to know which agent can perform which capability under which constraints.
MAIP v0.1 therefore includes a capability-discovery model intended to help agents identify available capabilities before delegating work.
Agent A
│
│ capability discovery
▼
┌─────────────────────────────┐
│ Agent Capability Document │
├─────────────────────────────┤
│ identity │
│ capabilities │
│ supported actions │
│ constraints │
│ interaction information │
└──────────────┬──────────────┘
│
▼
Agent B selected
│
▼
MAIP task request
The machine-readable capability definition is provided under /schema/.
A simplified multi-agent workflow can look like this:
Physician
│
▼
Doctor Agent
│
Clinical Intent
│
▼
MAIP
│
┌─────────────┼─────────────┐
▼ ▼ ▼
Laboratory Radiology Pharmacy
Agent Agent Agent
│ │ │
▼ ▼ ▼
Orders / Imaging Medication
Results Result Verification
│ │ │
└─────────────┼─────────────┘
▼
Shared Clinical Context
│
▼
Provenance / Audit
This illustrates the protocol's intended coordination model. It is not a claim that all of these workflow components are fully implemented by the current reference implementation.
The repository contains machine-readable schemas for the protocol's structured artifacts.
/schema/
├── message envelope
└── agent capability document
The canonical protocol definition is maintained in:
The schema files are intended to make the protocol implementable by systems that need machine-readable validation.
/
├── README.md — project overview
├── SPECIFICATION.md — MAIP v0.1 technical specification
├── CHANGELOG.md — version history
└── schema/ — JSON Schema definitions
AI-HOS is the reference implementation environment for MAIP.
AI-HOS is an AI-Native Hospital Operating System in which hospital workflows are designed around AI-driven coordination across clinical and operational departments.
MAIP itself is intended to remain implementation-independent. Its purpose is to define a protocol model that can be implemented by healthcare AI systems using different vendors, models, agent frameworks, and infrastructure.
Reference implementation:
The architecture and protocol design rationale are described in the accompanying AI-HOS Technical Whitepaper:
Ramadan, M. S. (2026). AI-HOS: An AI-Native Hospital Operating System with a Healthcare Agent Interaction Protocol for Context-Centric, Policy-Governed Clinical Workflow Orchestration. Technical Whitepaper v1.0. Sky Host Software.
The published record is available through Zenodo.
@techreport{ramadan2026aihos,
title = {AI-HOS: An AI-Native Hospital Operating System with a Healthcare Agent Interaction Protocol for Context-Centric, Policy-Governed Clinical Workflow Orchestration},
author = {Ramadan, Mohamed Salah},
institution = {Sky Host Software},
year = {2026},
type = {Technical Whitepaper},
version = {1.0},
doi = {10.5281/zenodo.22160266}
}MAIP v0.1 is a proposed specification under active design.
It is:
- a research and engineering specification
- open for implementation feedback
- designed for healthcare multi-agent coordination
- accompanied by machine-readable schemas
- associated with the AI-HOS reference implementation
It is not:
- an adopted industry standard
- a regulatory standard
- a claim of clinical deployment
- a claim of regulatory approval
- a claim of third-party adoption
Breaking changes may occur before a future v1.0 release.
MAIP v0.1 intentionally leaves several areas for continued protocol development.
See SPECIFICATION.md §10 for the current open issues and planned work toward v0.2.
Areas of ongoing protocol work include:
- retry and timeout semantics
- idempotency rules
- delegation authority
- context versioning
- interoperability profiles
- security and policy enforcement
- implementation feedback
- further evaluation and validation
The protocol will evolve through explicit specification changes rather than treating v0.1 as a frozen standard.
MAIP is an open specification under active development.
Issues and pull requests are welcome for:
- specification clarifications
- protocol semantics
- implementation feedback
- schema improvements
- interoperability considerations
- security and governance considerations
- proposed extensions
Before proposing a new protocol feature, please review the relevant sections of SPECIFICATION.md and the current open issues.
Specification text: Creative Commons Attribution 4.0 International (CC BY 4.0)
Schema files: MIT License
Eng. Mohamed Salah Ramadan
CEO & Lead Architect, Sky Host Software
For protocol discussions, implementation feedback, issues, and contributions, please use the GitHub repository's Issues and Pull Requests.
AI-HOS — AI-Native Hospital Operating System
https://m.salah.sky-host.pro/
AI-HOS Technical Whitepaper
https://doi.org/10.5281/zenodo.22160266
MAIP v0.1 · Proposed · Open Specification · Healthcare Multi-Agent Coordination