ChatGPT Agent 模式 2026：从对话到全自动执行

# 伪代码示例：Agent 任务规划
from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-5",
    instructions="You are an autonomous agent. Break down the user's request into actionable steps.",
    input="帮我调研当前主流的开源向量数据库，对比性能，生成报告",
    tools=[
        {"type": "web_search_preview"},
        {"type": "code_interpreter"},
        {"type": "file_search"},
    ],
    reasoning={"effort": "high"},
    # Agent 模式核心参数
    agent_config={
        "max_steps": 20,
        "auto_execute": True,
        "require_approval": "on_sensitive_action",
    }
)

for event in response:
    print(f"[{event.type}] {event.content}")

# Agent 反思循环的简化模型
class AgentLoop:
    def __init__(self, client, max_iterations=20):
        self.client = client
        self.max_iterations = max_iterations
        self.context = []
    
    def run(self, task: str):
        plan = self._plan(task)
        
        for step in plan:
            result = self._execute(step)
            evaluation = self._reflect(step, result)
            
            if evaluation.needs_revision:
                # 重新规划剩余步骤
                plan = self._replan(
                    task, 
                    completed=step, 
                    feedback=evaluation.feedback
                )
            
            self.context.append({
                "step": step,
                "result": result,
                "evaluation": evaluation
            })
        
        return self._synthesize(self.context)

import openai

client = openai.OpenAI()

# 启动 Agent 模式
agent = client.beta.agents.create(
    model="gpt-5",
    name="data-analyst",
    instructions="""
    你是一个数据分析专家。当收到分析任务时：
    1. 先搜索相关背景信息
    2. 编写 Python 代码进行数据处理
    3. 生成可视化图表
    4. 输出结构化分析报告
    """,
    tools=[
        {"type": "web_search_preview"},
        {"type": "code_interpreter", "config": {"max_runtime": 120}},
    ],
)

# 启动自主执行
thread = client.beta.agents.threads.create()
client.beta.agents.threads.messages.create(
    thread_id=thread.id,
    role="user",
    content="分析 2025 年全球 AI 芯片市场份额变化，预测 2026 趋势"
)

run = client.beta.agents.threads.runs.create(
    thread_id=thread.id,
    agent_id=agent.id,
    # 关键：启用自主执行模式
    agent_mode="autonomous",
    stream=True
)

# 流式接收 Agent 的执行过程
for event in run:
    if event.type == "tool_call":
        print(f"🔧 调用工具: {event.tool_name}")
    elif event.type == "tool_result":
        print(f"✅ 工具返回: {event.summary}")
    elif event.type == "reasoning":
        print(f"🧠 思考: {event.thought}")
    elif event.type == "message":
        print(f"💬 输出: {event.content}")

# 权限配置示例
agent_config = {
    "permissions": {
        # 只读操作：自动允许
        "web_search": "auto",
        "file_read": "auto",
        "code_execution": "auto",
        
        # 写操作：需要确认
        "file_write": "require_approval",
        "email_send": "require_approval",
        
        # 敏感操作：禁止
        "payment": "disabled",
        "account_deletion": "disabled",
    },
    
    # 资源限制
    "limits": {
        "max_steps": 50,
        "max_cost": 5.00,  # 美元
        "max_duration": 1800,  # 秒
    },
    
    # 审计日志
    "audit_log": True,
    "log_level": "detailed"
}

# 企业客服 Agent 配置
enterprise_agent = {
    "model": "gpt-5",
    "tools": [
        {"type": "function", "function": {
            "name": "query_order",
            "parameters": {"order_id": "string"}
        }},
        {"type": "function", "function": {
            "name": "process_refund",
            "parameters": {"order_id": "string", "amount": "number"}
        }},
    ],
    "permissions": {
        "query_order": "auto",
        "process_refund": "require_approval",
    }
}