Agentic Hospital — Implementation Report

1. System Overview

Agentic Hospital is a fully simulated AI-powered hospital built on Google ADK. It models a 500-bed academic medical centre with a multi-agent architecture where a root coordinator triages incoming patient queries and routes them to one of 40 specialist sub-agents. The system supports text conversations, medical image analysis, real-time web search for clinical guidelines, critical lab value monitoring, a live bed management simulation with ward visualisations, SQLAlchemy-backed persistence, BioMCP biomedical research tools, and coordinator-mediated inter-department specialist consults.

2. Architecture

2.1 System Block Diagram

2.2 Agent Hierarchy

hospital_coordinator  (root_agent — ADK entry point)
│
├── Tools (10):
│   get_patient_info · get_patient_encounter_history · check_critical_lab_values
│   triage_assessment · request_mdt_consultation · generate_treatment_plan
│   record_patient_encounter · generate_deterioration_alert
│   analyze_medical_image · web_search
│
└── Sub-Agents (39):
    ├── Medical Specialties (17)
    │   allergy_immunology · cardiology · endocrinology · gastroenterology
    │   general_medicine · gynecology · hematology · infectious_diseases
    │   nephrology · neurology · oncology · pediatrics · psychology
    │   pulmonology · rheumatology
    │
    ├── Surgical Specialties (9)
    │   cardiothoracic_surgery · colorectal_surgery · general_surgery
    │   neurosurgery · orthopedics · plastic_surgery · thoracic_surgery
    │   urology · vascular_surgery
    │
    ├── Diagnostic & Support Specialties (9)
    │   anesthesiology · critical_care · emergency_medicine · nuclear_medicine
    │   ophthalmology · pathology · physical_medicine_rehab · radiation_oncology
    │   radiology
    │
    └── Hospital Operations (4)
        hospital_admission · triage_nurse · pharmacy · discharge_planning

2.3 Patient Intake Flow (4-Phase + Admission) Updated: Sprints 1-4

2.4 Patient Flow (Updated with Sprints 1-4)

Patient arrives
      │
      ▼
hospital_coordinator (Phase 1: Registration)
   └─ get_patient_info + get_patient_encounter_history + check_critical_lab_values
      │
      ▼
triage_nurse_agent (Phase 2: Nurse Triage) [Sprint 1]
   ├─ calculate_esi_score() → ESI 1-5
   ├─ record_nurse_triage() → vitals, pain, allergies
   └─ assign_waiting_priority() → queue or resuscitation
      │
      ├─ ESI 1-2 → emergency_medicine_agent / critical_care_agent (immediate)
      └─ ESI 3-5 → hospital_coordinator (Phase 3: Routing)
                        │
                        ▼
                   specialist_agent
                      ├─ get_patient_info
                      ├─ get_lab_results / order_investigation() [Sprint 4]
                      ├─ record_vitals
                      ├─ check_drug_interactions
                      ├─ [2 specialty tools]
                      ├─ web_search
                      └─ generate_soap_note
                        │
                        ▼
                  If medication ordered:
                   pharmacy_agent [Sprint 2]
                      ├─ verify_medication_order()
                      ├─ medication_reconciliation()
                      └─ dispense_medication()
                        │
                        ▼
                  If admission:
                   hospital_admission_agent
                      └─ assign_bed()
                        │
                        ▼
                  If discharge:
                   discharge_planning_agent [Sprint 3]
                      ├─ check_discharge_criteria()
                      ├─ generate_discharge_summary()
                      ├─ pharmacy_agent.generate_tta_prescription()
                      ├─ send_gp_letter()
                      └─ arrange_community_services()

3. Agent Roster (39 sub-agents)

3.1 Root Coordinator

Coordinator tools (10):

3.2 Medical Specialties (17 agents)

3.3 Surgical Specialties (9 agents)

3.4 Diagnostic & Support Specialties (9 agents)

3.5 Hospital Operations (4 agents) Sprints 1-3

4. Tool Inventory

4.1 Common Tools — tools/common_tools.py

Available to every department agent (12 + 3 diagnostic functions):

4.2 Monitoring Tools — tools/monitoring_tools.py

4.3 Image Analysis Tool — tools/image_tools.py

Supported image types (13):

Input formats: File path, HTTPS URL, raw base64, or existing data URI.

4.4 Web Search Tools — tools/websearch_tools.py

Returns: {"status": "success"|"error"|"timeout", "query": ..., "result": ...}

4.5 Bed Management Tools — tools/bed_management_tools.py

Available to hospital_admission_agent only (8 functions):

4.6 Triage Tools — tools/triage_tools.py Sprint 1

Available to triage_nurse_agent (4 functions):

4.7 Pharmacy Tools — tools/pharmacy_tools.py Sprint 2

Available to pharmacy_agent (4 functions):

4.8 Discharge Planning Tools — tools/discharge_tools.py Sprint 3

Available to discharge_planning_agent (4 functions):

4.9 Diagnostic Ordering — tools/common_tools.py Sprint 4

Available to all agents (3 new functions):

4.10 Specialty Tools — tools/{specialty}_tools.py

35 files × 2 functions = 70 specialty tool functions total.

4.11 Knowledge Base Tool — tools/knowledge_base_tools.py Sprint 5

Available to triage_nurse, pharmacy, discharge_planning agents and progressively rolled out to all departments (1 function):

4.12 Authoritative Website Config — tools/website_config.py Sprint 5

Maps every department to its primary professional-society and guideline sources. Used by the make_department_web_search() factory in websearch_tools.py to bias browser-use queries toward evidence-based, peer-reviewed content instead of general health blogs.

4.13 Tool Count Summary

5. Patient & Lab Database

5.1 Patient Registry — _PATIENT_DB in common_tools.py

10 diverse mock patients (P001–P010) with realistic comorbidities, medications, and allergies:

5.2 Lab Database

Full panels available for all 10 patients including: CBC, BMP (electrolytes, renal, glucose), Lipid Panel, HbA1c, LFTs, TSH, coagulation (INR/PT), iron studies, hormone panels, troponin, BNP, urinalysis, blood cultures, and specialty panels.

5.3 Drug Interaction Database

In-memory lookup of known clinically significant drug-drug interactions checked before any new medication is recommended (_DRUG_INTERACTIONS in common_tools.py).

5.4 Drug Formulary

Dosing reference for common medications with weight- and organ-function-adjusted calculations used by calculate_medication_dose (_DRUG_FORMULARY in common_tools.py).

6. Hospital Infrastructure — Bed Management System

6.1 Overview

Added in the most recent implementation sprint. Provides a full simulation of hospital inpatient operations with real-time state tracking, ASCII ward maps, and a markdown hospital dashboard.

6.2 Ward Inventory

6.3 Bed States

6.4 Visualisation Examples

Hospital Dashboard (from get_hospital_dashboard()):

## 🏥 AGENTIC HOSPITAL — BED MANAGEMENT DASHBOARD
Date: 2026-02-20 14:35  |  Total Capacity: 325  |  Occupied: 215  |  Available: 78  |  Occupancy: 66%

| Ward                      | Capacity | 🔴 Occupied | 🟡 Cleaning | 🔧 Maint | 🟢 Available | Occ%    |
|---------------------------|----------|-------------|-------------|----------|--------------|---------|
| ICU                       |    10    |      6      |      1      |    1     |      2       |   60%   |
| Emergency                 |    15    |     10      |      2      |    0     |      3       |   67%   |
| Cardiology                |    20    |     14      |      2      |    0     |      4       |   70%   |
| ...

Ward Floor Plan (from get_ward_visualization("ICU")):

╔══════════════════════════════════════════════════════════════════╗
║  🏥 ICU Ward — Agentic Hospital                                  ║
║  Floor 2 | Nurse Station: ICU Central                            ║
╠══════════════════════════════════════════════════════════════════╣
║  Beds: 10 total | 🔴 6 occupied | 🟡 1 cleaning | 🟢 2 available ║
╠══════════════════════════════════════════════════════════════════╣
║  [ICU-01] 🔴 OCCUPIED   │ Robert Chen                           ║
║                          │ Dx: COPD exacerbation + CHF           ║
║                          │ Admitted: 2026-02-18 08:30            ║
╠══════════════════════════════════════════════════════════════════╣
║  [ICU-02] 🔴 OCCUPIED   │ George Thompson                       ║
║                          │ Dx: Decompensated HF + Persistent AFib║
╠══════════════════════════════════════════════════════════════════╣
║  [ICU-07] 🟡 CLEANING   │ Post-discharge cleaning — ready ~15min║
╠══════════════════════════════════════════════════════════════════╣
║  [ICU-08] 🟢 AVAILABLE  │ Ready for admission                   ║
╚══════════════════════════════════════════════════════════════════╝
  Occupancy: 6/10  |  Available now: 2

6.5 Admission Workflow (simulated)

Query: "Admit patient P007 to Cardiology for diabetic monitoring"
  ↓
coordinator → hospital_admission_agent
  ↓
1. get_patient_info("P007")             → Carlos Rivera, T1DM, allergic to amoxicillin
2. check_bed_availability("Cardiology") → 4 beds available
3. assign_bed("P007", "Cardiology", ...) → CARD-17 assigned
4. get_ward_visualization("Cardiology") → updated floor map shown

6.6 Waitlist Priority

Priority queue per ward:
  emergency (0) → immediate — life-threatening, ICU-level
  urgent    (1) → within 24h — significant illness
  routine   (2) → elective — stable or scheduled

Estimated wait: 30 min base + 45 min per patient ahead in queue

7. Clinical Reasoning Framework

Every department agent prompt embeds four shared constants from prompts/shared.py:

7.5 Chain of Expert Thought (CoET) — Knowledge Base System Sprint 5

Each department has a dedicated knowledge base file (knowledge_bases/{dept}_kb.py) containing structured clinical reasoning data. The query_chain_of_thought tool gives agents access to this at any point in a consultation via 6 query types:

7.6 MCP Integration — medical-mcp & FHIR R4 Sprint 5

7.6a medical-mcp — Clinical Knowledge Server

The mcps/medical_mcp_toolset.py module connects agents to the medical-mcp Node.js MCP server via ADK's McpToolset / stdio transport. This gives agents real-time access to FDA drug labels, RxNorm, PubMed, Google Scholar, clinical guidelines, and WHO statistics.

7.6b LangCare FHIR MCP — Real FHIR R4 Patient Records

The mcps/fhir_mcp_toolset.py module integrates the MIT-licensed langcare-mcp-fhir Go binary as a second MCP server. It connects to the public HAPI FHIR R4 test server (https://hapi.fhir.org/baseR4) — no authentication required. Configuration lives in mcps/fhir_config.yaml.

Agents can now read and write real FHIR R4 resources: Patient, Encounter, MedicationRequest, Condition, Observation, DiagnosticReport, CarePlan, and any other standard FHIR resource type.

7.7 BioMCP — Biomedical Research MCP Sprint 8

mcps/biomcp_mcp_toolset.py integrates the BioMCP Python binary as a third MCP server. It provides real-time access to 15+ biomedical databases (PubMed, ClinicalTrials.gov, ClinVar, gnomAD, OncoKB, OpenFDA FAERS, NCI CTS API, and more) through 36 structured tools.

Install: pip install biomcp-python  ·  Command: biomcp run (stdio)

7.8 SQLAlchemy DB Persistence Sprint 6

All critical clinical state is persisted in a SQLite database (agentic_hospital.db) via SQLAlchemy ORM. The DB is auto-created and seeded on first run through __init__.py → init_db().

7.9 Automated Test Suite Sprint 7

A comprehensive pytest suite lives in automated_test_debug_fix/ covering all tool layers.

7.10 Inter-Agent Specialist Consults Sprint 9

Department agents can now request peer specialist input during a consultation. The root coordinator automatically detects pending consults and routes to the target specialist before proceeding to prescription or discharge.

cardiology_agent → request_specialist_consult("P001","cardiology","hematology",
                       "Suspected HIT — switch anticoagulant?", urgency="urgent")
  ↓ writes ConsultRequest(status="pending") to DB
hospital_coordinator ← transfer_to_agent returns
  ↓ calls get_pending_consults("P001")  → [{target_agent:"hematology_agent", urgency:"urgent", ...}]
  ↓ calls transfer_to_agent("hematology_agent")
hematology_agent → reviews, provides opinion
  ↓ calls fulfill_specialist_consult("CONS-...", "hematology", response, recommendations)
    ↓ ConsultRequest(status="fulfilled") in DB
hospital_coordinator → continues to prescription / discharge

8. Prompt Architecture

Structure

Each department prompt follows this template:

You are Dr. {Name}AI, a {Department} Specialist at Agentic Hospital.

PERSONA & PHILOSOPHY:   [fellowship training, clinical philosophy]
EXPERTISE:              [15–20 specialty conditions/procedures]
CLINICAL APPROACH:
  ► KEY HISTORY QUESTIONS:    [department-specific intake questions]
  ► VALIDATED SCORING SYSTEMS:[5–8 validated clinical scores]
  ► EVIDENCE-BASED GUIDELINES:[5+ major society guideline citations]
  ► DIAGNOSTIC PITFALLS:      [5–8 specialty-specific cognitive traps]
EMERGENCY RED FLAGS:    [4–6 critical presentations requiring 911]

{_TOOL_PROTOCOL}
{_IMAGE_ANALYSIS_WORKFLOW}
{_CLINICAL_REASONING}
{_SAFETY_DISCLAIMER}

Prompt Files

9. File Structure

agentic_hospital/
├── __init__.py                         exports root_agent
├── agent.py                            root coordinator + 39 sub-agent registration
├── config.py                           runtime configuration (model, env)
│
├── departments/                        39 agent definitions
│   ├── allergy_immunology.py
│   ├── anesthesiology.py
│   ├── cardiology.py
│   ├── cardiothoracic_surgery.py
│   ├── colorectal_surgery.py
│   ├── critical_care.py
│   ├── dermatology.py
│   ├── discharge_planning.py           ← discharge coordinator (Sprint 3)
│   ├── emergency_medicine.py
│   ├── endocrinology.py
│   ├── ent.py
│   ├── gastroenterology.py
│   ├── general_medicine.py
│   ├── general_surgery.py
│   ├── gynecology.py
│   ├── hematology.py
│   ├── hospital_admission.py           ← bed management agent
│   ├── infectious_diseases.py
│   ├── nephrology.py
│   ├── neurology.py
│   ├── neurosurgery.py
│   ├── nuclear_medicine.py
│   ├── oncology.py
│   ├── ophthalmology.py
│   ├── orthopedics.py
│   ├── pathology.py
│   ├── pediatrics.py
│   ├── pharmacy.py                     ← clinical pharmacist (Sprint 2)
│   ├── physical_medicine_rehab.py
│   ├── plastic_surgery.py
│   ├── psychology.py
│   ├── pulmonology.py
│   ├── radiation_oncology.py
│   ├── radiology.py
│   ├── rheumatology.py
│   ├── thoracic_surgery.py
│   ├── triage_nurse.py                 ← ESI v4 triage nurse (Sprint 1)
│   ├── urology.py
│   └── vascular_surgery.py
│
├── knowledge_bases/                    38 files — CoET knowledge base system (Sprint 5)
│   ├── base_reasoning.py               SPECIALTY_REGISTRY + process_coet_query() engine
│   ├── {dept}_kb.py × 35              35 medical/surgical/diagnostic specialty KBs
│   └── discharge_planning_kb.py        discharge planning KB (36th specialty KB)
│
├── mcps/                               MCP server integrations (Sprint 5)
│   ├── medical_mcp.py                  Python API wrappers (fallback)
│   ├── medical_mcp_toolset.py          ADK McpToolset instances — 11 toolsets (medical-mcp)
│   ├── fhir_config.yaml                HAPI FHIR R4 server configuration (no auth)
│   └── fhir_mcp_toolset.py             LangCare FHIR MCP toolsets — 4 toolsets (MIT)
│
├── prompts/                            42 prompt files
│   ├── shared.py                       _TOOL_PROTOCOL, _CLINICAL_REASONING, _CHAIN_OF_EXPERT_THOUGHT
│   ├── coordinator.py
│   ├── hospital_admission.py
│   ├── triage_nurse.py
│   ├── pharmacy.py
│   ├── discharge_planning.py
│   ├── department_prompts.py           backward-compat re-export shim
│   └── {dept}.py × 35
│
└── tools/                             46 tool files, 114 functions total
    ├── common_tools.py                 15 functions (12 core + 3 diagnostic ordering)
    ├── monitoring_tools.py             2 critical alert functions
    ├── image_tools.py                  1 multimodal analysis function
    ├── websearch_tools.py              1 browser-use search + make_department_web_search()
    ├── website_config.py               4 utility functions — authoritative site config per dept (Sprint 5)
    ├── bed_management_tools.py         8 bed management functions + _BED_DB
    ├── triage_tools.py                 4 ESI v4 triage functions
    ├── pharmacy_tools.py               4 medication verification/dispensing functions
    ├── discharge_tools.py              4 discharge planning functions
    ├── knowledge_base_tools.py         1 CoET query function (query_chain_of_thought)
    └── {dept}_tools.py × 35           70 specialty functions (2 per department)

10. Dependencies

Environment variables required (.env):

GOOGLE_API_KEY=your_key_here

Browser dependency:

playwright install chromium   # required for web_search tool

11. Running the System

# Install dependencies
pip install -r requirements.txt
playwright install chromium

# Launch ADK web UI
adk web

# Opens at http://localhost:8000
# Select "hospital_coordinator" from the agent dropdown

Example Interactions

12. What's Next