Agentic Hospital: Agentifying Hospitals for AI-Native Clinical and Operational Applications

📄 Paper • 🏥 GitHub

Overview

Agentic Hospital introduces a new paradigm that models the hospital as an accessible socio-technical environment for autonomous agents. Unlike prior work that primarily emphasizes role simulation or medical interaction, Agentic Hospital enables agents to interact with simulated human roles, access and manipulate hospital information systems, retrieve evidence from local and external knowledge sources, invoke computational tools, and act under explicit governance constraints.

Under this formulation, an Agentic Hospital is a hospital-as-an-environment in which autonomous agents can directly perceive, query, and act on hospital resources through standardized, actionable interfaces — including human actors (e.g., doctors, nurses, patients, administrators) and digital-operational infrastructure (e.g., HIS/EMR-like systems, resource views, scheduling pipelines, and management tools).

Key Contributions

• Agentic Hospital: A new conceptualization of hospital intelligence in which autonomous agents can directly operate over both human and digital hospital environments through standardized interfaces.
• Concrete Implementation Framework: A framework that combines role simulation, mocked HIS, structured data views, evidence-based retrieval, peripheral management systems, and secure computational tools into a unified hospital-scale agent environment.
• Expert-Grounded Evaluation Protocol: An evaluation protocol that benchmarks both clinical and operational hospital tasks under structured rubrics, enabling systematic study of model capabilities in realistic hospital settings.

Framework

The framework consists of five core components:

Component	Description
Agent Layer	Autonomous agents that play hospital roles and perform planning, reasoning, and action selection
Environment Layer	Human environment (role simulation) + Digital-operational environment (HIS, structured data, workflows)
Interface and Capability Layer	Standardized interfaces (CLI commands, APIs, structured data views) that make the environment directly actionable
Governance and Evaluation Layer	Permission control, safe execution, logging, auditing, and performance assessment

Implementation

The implementation consists of six tightly connected parts:

1. Role Simulation — Large-scale role simulation for human actors in the hospital, covering both clinical roles (doctors, nurses, pharmacists, patients, etc.) and administrative-operational roles (billing staff, department administrators, hospital managers).

2. Mocked HIS — A mocked but actionable HIS/HER-like system with structured tables, agent-facing interfaces (CLI commands, APIs), and structured action schemas.

3. Structured Data Views — Views over local hospital cases and records that transform low-level hospital records into higher-level structured representations agents can directly consume.

4. Evidence-Based Search — A retrieval module built over both local and external knowledge sources (clinical literature, practice guidelines, drug references, institutional policies).

5. Management and Peripheral Systems — Financial/settlement, pharmacy, logistics, administrative oversight, and policy/compliance systems that capture the broader institutional ecosystem.

6. Execution Tools — Python interpreters, optimization solvers, structured data processing, evidence retrieval, medication safety checkers, clinical risk analyzers, and more.

Simulated Roles

Category	Roles
Clinical	Doctor, Nurse, Patient, Standardized Patient, Pharmacist, Medical Technician
Administrative-Operational	Registration Staff, Billing and Settlement Staff, Department Administrator, Hospital Manager, Logistics and Inventory Staff, Pharmaceutical Representative

Mocked HIS Modules

Module	Representative APIs	Supported Tasks
Patient and EMR	get_patient_record, update_diagnosis	Patient profiling, diagnostic support, longitudinal record review
Laboratory and Imaging	query_lab_results, create_lab_order	Evidence collection, automated planning
Medication and Pharmacy	create_medication_order, check_drug_inventory	Medication review, prescription support
Scheduling and Info	schedule_appointment, check_bed_availability	Appointment scheduling, bed allocation
Billing and Settlement	review_billing_summary, check_settlement_status	Cost-aware planning, reimbursement support

Evaluation

The evaluation protocol features:

• Expert-grounded rubric-based evaluation spanning both clinical and operational domains
• Structured rubrics with axes including correctness, safety, completeness, compliance, and efficiency
• Task diversity: expert annotations, clinical and operational hospital tasks
• Benchmark quality: annotator record verification and consistency checks throughout the annotation process

Comparison with Prior Work

	AI Hospital	Agent Hospital	Agentic Hospital (Ours)
Case scope	Single-agent diagnosis	Closed-loop hospital rooms	Actionable hospital environment for autonomous agents
Environment	History-only interactions	Simulated hospital workflows	Human actors + hospital systems + tools + governance
System access	Mostly implicit	Partially simulated	Directly actionable via HIS/EMR/CLI/APIs
Agent ability	Dialogue and diagnosis	Planning and self-evolution	Perceive, reason, call tools, access data, and execute

Citation

@article{agentic_hospital2025,
  title={Agentic Hospital: Agentifying Hospitals for AI-Native Clinical and Operational Applications},
  author={Wenyuan Gu and Hongyuan Zha and Benyou Wang},
  year={2025}
}

Contact

• Wenyuan Gu, Hongyuan Zha, Benyou Wang
• The Chinese University of Hong Kong, Shenzhen
• Email: wangby@cuhk.edu.cn