OpenAI Agents SDK — Multi-Agent Workflows in Python

What it is

OpenAI Agents SDK is the official lightweight Python framework from OpenAI for building multi-agent workflows. It provides core primitives: Agents (LLMs with instructions and tools), Handoffs (delegation between specialized agents), Guardrails (input/output validation), and Tracing (execution observability). The SDK is designed to be minimal yet production-ready.

The SDK targets Python developers building multi-agent systems with OpenAI models. It works with GPT-4o, o3, and any OpenAI-compatible endpoint, making it suitable for both cloud and self-hosted model deployments.

How it saves time or tokens

The SDK handles agent orchestration, context passing, and error recovery that you would otherwise build from scratch. Handoffs let specialized agents handle subtasks (billing, technical support, research) without routing logic in your application code. Guardrails validate inputs and outputs before and after agent runs, catching issues early and reducing wasted API calls.

How to use

1. Install the SDK: pip install openai-agents.
2. Define agents with instructions and tools, and configure handoffs between them.
3. Run the workflow with Runner.run_sync(agent, input) or the async equivalent.

Example

from agents import Agent, Runner

# Define specialized agents
billing_agent = Agent(
    name='Billing',
    instructions='Handle billing questions and refunds.'
)

technical_agent = Agent(
    name='Technical',
    instructions='Handle technical support questions.'
)

# Triage agent delegates to specialists
triage = Agent(
    name='Triage',
    instructions='Route the user to the right agent.',
    handoffs=[billing_agent, technical_agent]
)

# Run the workflow
result = Runner.run_sync(triage, 'I need a refund for my last order.')
print(result.final_output)

Related on TokRepo

• Multi-Agent Frameworks -- compare multi-agent orchestration frameworks side by side
• AI Tools for Agents -- explore agent-building tools and frameworks

Common pitfalls

• The SDK is optimized for OpenAI models; using non-OpenAI endpoints may require adjusting prompt formats for handoff instructions.
• Guardrails run additional LLM calls for validation, which adds latency and cost; use them selectively on high-risk inputs.
• Tracing generates detailed execution logs that can be verbose; configure log levels appropriately for production deployments.