Learning: OpenAI Agent SDK - Agent Handoffs Pattern

https://www.pexels.com/photo/people-passing-a-baton-12585936/

Following these blogs Learning: OpenAI Agent SDK and Learning: OpenAI Agent SDK - Conditional tool enabling patterns & Run context, this blog outlines the implementation of agent handoff patterns in agentic systems, demonstrating how a triage agent can intelligently transfer conversations to specialized medical department agents using the OpenAI Agent SDK with Azure AI Foundry integration.

All the code examples discussed in this series are available in our GitHub repository. We've chosen the healthcare domain as our focus area because it provides clear, practical examples of how agentic patterns can solve real-world problems while demonstrating the importance of reliability and accuracy in AI systems.

Overview

Agent handoff patterns enable seamless transfer of conversation control between different specialized agents based on user needs or medical conditions. This approach is crucial for building conversational systems where different agents excel at specific medical domains or patient care scenarios.

In medical triage contexts, handoff enables routing patients to appropriate medical departments based on their symptoms or conditions. For specialized care, this pattern allows transferring conversations to domain-specific medical specialists who can provide more accurate and detailed assistance.

Conversation continuity is maintained as the handoff preserves the entire conversation history when transferring between agents. Additionally, handoff facilitates progressive specialization where patients can be initially triaged and then seamlessly transferred to increasingly specialized care providers as needed.

Key Concepts

Triage Agent

The initial point of contact that analyzes patient descriptions and medical conditions to determine the most appropriate specialist department for handoff.

Specialist Department Agents

Individual agents representing different medical departments (Emergency, Surgery, ICU, Cardiology, etc.) with specialized knowledge and care protocols for their respective domains.

Agent Handoff

The process by which the triage agent transfers conversation control to a specialist agent, maintaining conversation history and context throughout the transition.

Architecture

The following diagram illustrates the agent handoff architecture:

Flow Description

1. Patient Input: Description of medical condition, symptoms, or health concerns
2. Triage Agent: Analyzes the patient's situation and determines the most appropriate medical department
3. Specialist Department Agents: Take over the conversation to provide specialized medical guidance and care
4. Specialized Response: Deliver domain-specific medical advice, treatment recommendations, or care instructions

Best Practices

1. Clear Handoff Instructions

Ensure the triage agent has clear instructions for routing decisions:

```python
instructions=(
    "Handoff to the appropriate agent based on the medical condition of "
    "the patient. Analyze symptoms carefully and route to the most "
    "suitable specialist department."
)```

2. Specialized Agent Instructions

Each specialist agent should have domain-specific instructions:

```python
emergency_department_agent = Agent(
    name="emergency_department_agent",
    instructions=(
        "Handle emergency medical situations promptly and efficiently. "
        "Triage patients based on the severity of their conditions and "
        "provide immediate care guidance."
    ),
    model=llm_model,
  )```

Usage

Running the Example

```bash
# Activate virtual environment
source .venv/bin/activate

# Run the conditional tool enabling example
python -m openai_agent.agentic_patterns.tool_conditional
```

Demo (please view the video in full screen and it does not have audio)