AI-HOS: An AI-Native Hospital Operating System with a Healthcare Agent Interaction Protocol for Context-Centric, Policy-Governed Clinical Workflow Orchestration

Abstract

Modern hospital information systems (HIS) and electronic health records (EHR) are designed around application modules, forms, and database records accessed through CRUD APIs by human users. This paper argues that this application-centric paradigm is structurally insufficient for a new class of healthcare workflows in which heterogeneous artificial intelligence agents must coordinate over shared patient context, execute authorized actions under policy constraints, and remain auditable end-to-end. We propose AI-HOS, an AI-native hospital operating system architecture in which the hospital is modeled as a context-centric, policy-governed, agentic operating environment, and MAIP (Medical Agent Interaction Protocol), a healthcare-specific agent interaction protocol that complements — rather than replaces — existing standards such as FHIR, SMART on FHIR, CDS Hooks, A2A, and MCP. We describe the AI-HOS eleven-layer reference architecture, the MAIP message envelope and task lifecycle, a clinical safety model with action classes and autonomy levels, a provenance and audit chain, and a proposed evaluation framework with eight formal metrics and five baseline comparisons. We deliberately do not report experimental results: this is a position and architecture whitepaper, not a peer-reviewed research paper. A v2 paper with experimental evaluation will follow the first pilot deployment. The contributions are: (1) the AI-HOS architecture, (2) the MAIP protocol proposal, and (3) a measurement framework for agentic hospital systems.