OpenAI Agents SDK Python教程：多Agent协作开发指南

日期：2026-08-14 | 分类：AI工具教程

项目简介

OpenAI Agents SDK是OpenAI官方开源的Python多Agent开发框架，支持Agent创建、工具调用、Agent间Handoff移交、Guardrails安全检查、Tracing全链路追踪。

安装

pip install openai-agents

# 设置API Key
export OPENAI_API_KEY="sk-your-key"

创建第一个Agent

from agents import Agent, Runner

# 定义Agent
agent = Agent(
    name="代码助手",
    instructions="你是一个Python代码专家，帮助用户编写和调试代码",
    model="gpt-5.6"
)

# 运行Agent
result = Runner.run_sync(agent, "帮我写一个快速排序")
print(result.final_output)

定义工具

from agents import function_tool

@function_tool
def search_code(query: str) -> str:
    """搜索代码库"""
    # 实际搜索逻辑
    return f"找到关于{query}的3个结果"

@function_tool
def run_code(code: str) -> str:
    """执行代码并返回结果"""
    import subprocess
    result = subprocess.run(["python", "-c", code], capture_output=True)
    return result.stdout.decode()

agent = Agent(
    name="开发助手",
    instructions="帮助用户搜索和运行代码",
    tools=[search_code, run_code]
)

多Agent Handoff移交

from agents import Agent, Runner

# 前端Agent
frontend_agent = Agent(
    name="前端专家",
    instructions="专精React/Vue前端开发",
)

# 后端Agent
backend_agent = Agent(
    name="后端专家",
    instructions="专精Python/Go后端开发",
)

# 路由Agent
triage_agent = Agent(
    name="路由助手",
    instructions="根据问题类型移交给对应专家",
    handoffs=[frontend_agent, backend_agent]
)

result = Runner.run_sync(triage_agent, "帮我写一个REST API")
print(result.final_output)

Guardrails安全护栏

from agents import GuardrailFunctionOutput, Agent, Runner

async def safety_check(ctx, agent, input):
    is_safe = "恶意" not in str(input)
    return GuardrailFunctionOutput(
        output_info={"safe": is_safe},
        tripwire_triggered=not is_safe
    )

agent = Agent(
    name="安全助手",
    instructions="回答用户问题",
    input_guardrails=[safety_check]
)

Tracing追踪

from agents import trace

with trace("多Agent工作流"):
    result1 = Runner.run_sync(agent1, "分析需求")
    result2 = Runner.run_sync(agent2, result1.final_output)
    # 在platform.openai.com查看完整追踪

相关阅读

📚 常见问题

OpenAI Agents SDK Python是什么？

OpenAI Agents SDK是OpenAI官方开源的Python多Agent开发框架，支持Agent创建、工具调用、Agent间Handoff移交、Guardrails安全检查、Tracing全链路追踪。

如何上手OpenAI Agents SDK Python？

2026 Agent框架选型指南