OpenAI Agents SDK — Multi-Agent Workflows in Python

Core Concepts

Agents

Agents are LLMs configured with instructions and tools:

agent = Agent(
    name="Research Agent",
    instructions="You research topics thoroughly.",
    tools=[search_tool, wikipedia_tool],
)

Handoffs

Agents can delegate to specialized agents:

triage = Agent(
    name="Triage",
    handoffs=[billing_agent, technical_agent, general_agent],
)

Guardrails

Validate inputs/outputs before and after agent runs:

@input_guardrail
async def check_topic(ctx, agent, input):
    # reject off-topic requests
    ...

Tracing

Built-in tracing for debugging multi-agent flows — visualize handoff chains, tool calls, and token usage.

FAQ

Q: What is OpenAI Agents SDK? A: The official OpenAI Python framework for multi-agent workflows with handoffs, tool calling, guardrails, and tracing. 20K+ GitHub stars.

Q: How is it different from LangChain or CrewAI? A: It's minimal by design — focused primitives (agents, handoffs, guardrails) without a large dependency tree. Built and maintained by OpenAI.