LangGraph Healthcare AI Agent Suite

Stateful Symptom Triage, Human-in-the-Loop Doctor Verification & Multi-Agent Clinical Research Orchestrator

An enterprise-grade agentic workspace applying LangGraph and LangChain to healthcare assistant design patterns. Features three distinct architectural clinical patterns: basic symptom triage conversation, human-in-the-loop clinical decision support with manual doctor overrides, and supervisor-supervised multi-agent systems for medical risk analysis and research report generation.

---

How It Works

This repository contains independent modules built on top of LangGraph's state machine execution model. Each module is contextualized around a critical healthcare application, demonstrating how state management, loops, and human-in-the-loop workflows operate in high-integrity domains.

View Healthcare AI Agent Suite Flow Diagram

graph TD
    classDef main fill:#1C3C3C,stroke:#333,stroke-width:1px,color:#fff;
    classDef chatbot fill:#3776AB,stroke:#333,stroke-width:1px,color:#fff;
    classDef human fill:#F55036,stroke:#333,stroke-width:1px,color:#fff;
    classDef multi fill:#008080,stroke:#333,stroke-width:1px,color:#fff;
    classDef tool fill:#f4f4f4,stroke:#333,stroke-width:1px,color:#333;

    Workspace([User Clinical Request]) --> Route{Select Workflow}
    
    %% Pattern 1: Triage Bot
    Route -->|Symptom Triage| P1[1-BasicChatBot]:::chatbot
    P1 --> ChatbotNode["Triage Bot Node"]:::chatbot
    ChatbotNode -->|Triage Advice| Output1([Patient Instructions])

    %% Pattern 2: HIL Clinical Assistant
    Route -->|Doctor Approval| P2[Human Assistance]:::human
    P2 --> HILChat["Clinical Assistant Node"]:::human
    HILChat -->|Conditional Edge| HILTools[Tool Node]:::tool
    HILTools -->|TavilySearch| WebSearch["Medical Database Search"]:::tool
    HILTools -->|human_assistance| Interrupt["Doctor Override / Interrupt"]:::human
    Interrupt -->|Doctor Approval/Edits| ResumedChat[Resume Graph Execution]:::human

    %% Pattern 3: Clinical Supervisor
    Route -->|Clinical Multi-Agent| P3[Multi-Agent System]:::multi
    P3 --> Sup["Medical Supervisor Node"]:::human
    Sup --> Router{"Supervisor Router"}:::main
    Router -->|assign research| ResAgent["Medical Researcher Node"]:::multi
    Router -->|assign analysis| AnaAgent["Clinical Analyst Node"]:::multi
    Router -->|assign writing| WritAgent["Medical Writer Node"]:::multi
    Router -->|DONE| End([Compiled Case Brief])

    ResAgent -->|return research_data| Sup
    AnaAgent -->|return analysis| Sup
    WritAgent -->|return final_report| Sup

Loading

Clinical Safety Guardrails & Verification Architecture

Because clinical safety and accurate guidance are the primary objectives of this agent suite, a multi-tiered safety guardrails architecture is implemented across the graph designs:

graph TD
    classDef safety fill:#F55036,stroke:#333,stroke-width:1px,color:#fff;
    classDef agent fill:#3776AB,stroke:#333,stroke-width:1px,color:#fff;
    classDef tool fill:#f4f4f4,stroke:#333,stroke-width:1px,color:#333;

    Input([Clinical Input]) --> Triage[Symptom Triage Gate]:::safety
    Triage -->|Check Scope Constraints| Assistant[Clinical Assistant Node]:::agent
    
    Assistant -->|Generate Advice/Prescription| DoctorCheck{Human Doctor Review Gate}:::safety
    DoctorCheck -->|Interrupt State| ReviewNode[State Interrupt & Resume]:::tool
    ReviewNode -->|Clinical Override/Signature| Output[Patient Output]:::agent
    
    subgraph Multi-Agent Workspace
        Supervisor[Chief Medical Supervisor]:::safety -->|Handoff Policy Control| WorkerAgents[Specialized Worker Nodes]:::agent
    end

Loading

🛡️ Safety Rail Design Mappings

1. Symptom Triage Gate (Input Guardrails)

   • Mechanism: System prompts strictly forbid the triage chatbot from diagnosing conditions or prescribing medications.
   • Boundary: Limits assistant behavior to query collection and clinic routing recommendations.

2. Doctor Verification Gate (Human-in-the-Loop Resumption)

   • Mechanism: Suspends the thread execution using LangGraph interrupt when medical prescriptions or specific diagnosis labels are detected.
   • Boundary: The state is serialized to a persistent database checkpoint, preventing user disclosure until a licensed medical professional reviews and reschedules the state.

3. Supervisor Handoff Policy Control (Structured Delegation)

   • Mechanism: The supervisor enforces separation of concerns by reading custom variables (has_research, has_analysis, has_report).
   • Boundary: Prevents downstream agents from executing tasks out-of-order or accessing tools outside their clinical scope.

---

Key Features

• Clinical Symptom Triage — Conversational flows for gathering patient symptoms, history, and concerns using groq:llama-3.3-70b-versatile inside a LangGraph StateGraph.
• Human-in-the-Loop Verification — Ensures patient safety by interrupting execution when prescription or diagnostic outputs are generated, allowing a licensed doctor to review, override, and resume the graph run.
• Hierarchical Medical Supervision — Coordinated team of specialized agents led by a Chief Medical Supervisor who evaluates variables dynamically to route research, data analysis, and documentation.
• Fact-Based Medical Grounding — Integration of Tavily search engines to retrieve peer-reviewed medical publications, clinical guidelines, and drug database specs.
• State Persistence & Checkpointers — Full context-aware memory retention for conversational histories and doctor overrides via MemorySaver.

Tech Stack

Layer	Technology	Description
State Orchestration		Manages patient state, conditional clinical paths, node interrupts, and checkpointer states.
Agent Abstraction		Decoupled LLM schemas, doctor-override tool decorators, clinical prompts, and unified chat initialization.
Inference Engine		High-speed clinical reasoning using Groq Cloud Llama-3.3-70B model.
Grounding Engine		Queries medical journals, drug databases, and clinical trials for context-aware grounding.
Clinical Persistence		Enforces hard state checkpoints to enable pauses for clinical doctor override and safety checks.

Project Structure

.
├── 1-BasiChatBot/
│   └── 1-basicchatbot.ipynb      # Basic patient triage and dialogue graph
├── HumanAssitance/
│   └── HumaninLoop.ipynb         # Doctor-in-the-loop safety interrupt & review node
├── MultiAgents/
│   └── Agents.ipynb              # Medical Supervisor, Researcher, Analyst & Writer team
├── main.py                       # Global entrypoint script
├── pyproject.toml                # UV tool configuration & dependency specifications
├── requirements.txt              # Standard requirements file
├── .env                          # API secrets (GROQ & TAVILY)
└── README.md                     # Healthcare suite documentation

Getting Started

Prerequisites

• Python 3.13+
• A Groq API Key (get one from console.groq.com)
• A Tavily API Key (get one from tavily.com)

Installation

# Clone the repository
git clone https://github.com/your-username/langgraph-healthcare-agents.git
cd langgraph-healthcare-agents

# Create and activate a virtual environment
python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

# Install dependencies using uv (recommended)
uv pip install -r requirements.txt
# Or using standard pip:
pip install -r requirements.txt

Configuration

Create a .env file in the root of the project to configure your credentials:

GROQ_API_KEY=your_groq_api_key_here
TAVILY_API_KEY=your_tavily_api_key_here

---

The Workflows

1. Basic Patient Triage Assistant

Located in 1-basicchatbot.ipynb, this workflow configures a symptom collection and triage chat bot. The conversational agent prompts patients for symptom profiles, timeline, and intensity before offering routing suggestions (e.g., advising ER vs. primary care clinic visits).

2. Human-in-the-Loop Clinical Verification

Located in HumaninLoop.ipynb, this workflow implements strict human oversight safeguards. When the AI proposes diagnostic summaries or treatment suggestions, the system calls a human_assistance tool which registers an interrupt state.

graph TD
    classDef chatbot fill:#3776AB,stroke:#333,stroke-width:1px,color:#fff;
    classDef human fill:#F55036,stroke:#333,stroke-width:1px,color:#fff;
    classDef tool fill:#f4f4f4,stroke:#333,stroke-width:1px,color:#333;

    Start([Patient Input]) --> Assistant[Assistant Node]:::chatbot
    Assistant -->|Decide Diagnostic Suggestion| ToolNode[Tool Execution Node]:::tool
    ToolNode -->|Require Doctor Check| Interrupt[State Interrupt]:::human
    Interrupt -->|Doctor Approval/Correction| Resume[Resume Graph Execution]:::human
    Resume --> Output([Diagnostic Report Delivered])

Loading

Caution

Safety Critical Boundary No diagnosis, prescription, or clinical guidance is ever surfaced to the user output before state resumption. A licensed practitioner must approve the clinical output.

3. Multi-Agent Medical Supervisor Team

Located in Agents.ipynb, this notebook implements a hierarchical clinical workspace. A Chief Medical Supervisor coordinates a team of specialized agents, dynamically updating routing based on case state variables.

graph TD
    classDef supervisor fill:#F55036,stroke:#333,stroke-width:1px,color:#fff;
    classDef worker fill:#3776AB,stroke:#333,stroke-width:1px,color:#fff;
    classDef router fill:#1C3C3C,stroke:#333,stroke-width:1px,color:#fff;

    Case([Medical Case Input]) --> Sup[Chief Medical Supervisor]:::supervisor
    Sup -->|Determine Phase| Router{State Router}:::router
    
    Router -->|gather findings| Res["Medical Researcher Agent"]:::worker
    Router -->|evaluate safety| Ana["Clinical Analyst Agent"]:::worker
    Router -->|write case brief| Writ["Medical Writer Agent"]:::worker
    Router -->|all complete| End([Final Case Brief Compiled])

    Res -->|adds research_data| Sup
    Ana -->|adds analysis| Sup
    Writ -->|adds final_report| Sup

Loading

Note

SupervisorState Definition The state of the clinical supervisor model is tracked using custom schema definitions:

class SupervisorState(MessagesState):
    next_agent: str      # Tracks routing targets (researcher/analyst/writer/end)
    research_data: str   # Gathers clinical study results and literature checks
    analysis: str        # Evaluates drug interaction risks and diagnostic flags
    final_report: str    # Houses the finalized case brief compiled by the writer
    task_complete: bool  # Signal flag to trigger END
    current_task: str    # Stores original patient case file metadata

---

Customization

The medical agent suite is designed to be highly modular and extensible. You can adapt it to any clinical domain:

Component	Target File	Modification Details
Change LLM Model	pyproject.toml / Jupyter Notebooks	Modify init_chat_model("groq:llama-3.3-70b-versatile") parameters in notebooks to target different clinical fine-tunes.
Alter Supervisor Logic	Agents.ipynb	Edit prompt templates inside create_supervisor_chain to enforce strict compliance guidelines (e.g. HIPAA standards).
Add Specialized Agents	Agents.ipynb	Add new nodes (e.g. pharmacist_agent for drug interactions) and update router Literal endpoints.
Configure Local DB Checkpointing	HumaninLoop.ipynb	Replace in-memory checkpointer (MemorySaver) with persistent SQLite or Postgres savers for production clinical session tracking.

---

Acknowledgments

• Model Context Protocol (MCP) for standardizing application-tool interaction.
• LangGraph Documentation for state graph architecture and patterns.
• Groq Cloud for high-performance Llama 3.3 model inference.

---