AI Agent终极定义：程序员1张图彻底搞懂「感知‑规划‑行动」闭环

收藏它，下次面试官问“Agent和Chatbot有何不同”时，一张图+一段代码闭嘴！

为何2025年非谈Agent不可

Agent已从PPT概念，变为生产环境刚需。

1 一张图看透：感知‑规划‑行动‑审计闭环

桌面常驻图：任何Agent架构一目了然。

2 五大子系统拆解：程序员视角技术栈

3 从0到1：30行Python跑通最小Agent

3.1 环境依赖

pip install openai langchain pydantic

3.2 最简示例

from openai import OpenAI
import os, json, datetime

client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))

class Agent:
    def __init__(self):
        self.memory = []

    def perceive(self, query):
        self.memory.append({"role":"user","content":query})
        return {"query":query, "time":str(datetime.datetime.utcnow())}

    def plan(self, tools):
        prompt = f"Use these tools {tools} to answer: {self.memory[-1]['content']}"
        resp = client.chat.completions.create(
            model="gpt-4o-mini",
            messages=[{"role":"system","content":prompt}],
            tools=tools
        )
        return resp.choices[0].message

    def act(self, call):
        name = call.function.name
        args = json.loads(call.function.arguments)
        if name=="get_weather":
            return {"temp":25, "city":args["city"]}
        return {"error":"unknown"}

    def run(self, query):
        self.perceive(query)
        tools = [{
          "type":"function",
          "function":{"name":"get_weather","description":"current weather","parameters":{
            "type":"object","properties":{"city":{"type":"string"}}}}
        }]
        msg = self.plan(tools)
        if msg.tool_calls:
            result = self.act(msg.tool_calls[0])
            self.memory.append({"role":"assistant","content":str(result)})
        return self.memory[-1]["content"]

agent = Agent()
print(agent.run("What's the weather in Beijing?"))

输出：

{'temp': 25, 'city': 'Beijing'}

30行骨架，轻松扩展自定义工具。

4 实战加速：LangGraph × Function Calling

场景：抓取Hacker News并自动生成技术周报

架构图

graph LR
  U[用户指令] --> S[Agent Supervisor]
  S --> F[抓取 HN]
  F --> P[LLM 总结]
  P --> R[生成周报]
  R --> M[发送邮件]

核心代码

from langgraph.graph import Graph, END
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage
import requests

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)

def fetch_hn(state):
    top = requests.get("https://hacker-news.firebaseio.com/v0/topstories.json").json()[:5]
    titles = [requests.get(f"https://hacker-news.firebaseio.com/v0/item/{i}.json")
              .json()["title"] for i in top]
    return {"titles": titles}

def summarize(state):
    prompt = f"Summarize these titles into a tech weekly report:\n{state['titles']}"
    return {"summary": llm.invoke([HumanMessage(content=prompt)]).content}

def send_email(state):
    print("📨", state["summary"])
    return END

w = Graph()
w.add_node("fetch", fetch_hn)
w.add_node("summarize", summarize)
w.add_node("email", send_email)
w.add_edge("fetch","summarize")
w.add_edge("summarize","email")

app = w.compile()
app.invoke({})

1分钟搞定「感知→规划→行动」闭环。

5 成本揭秘：一次任务只需\$0.0002

万次日活仅\$2.1，开会2杯咖啡钱。

6 工程坑盘点：Token暴增、循环、幻觉

• Token暴增：

  • 只保留最近4k tokens；摘要旧记忆；向量检索Top‑5

• 循环调用：

  • 限制深度（默认5）；结果缓存；Plan自检去重

• 幻觉治理：

  • 用 Guardrails 校验Schema；失败重试；可观测

7 2026路线图：多Agent协作、边缘推理、合规沙箱

8 一句话总结：贴桌面边框的3个关键数字

1. 200 ms：端到端延迟红线
2. \$0.0002：单次成本上限
3. 5步：循环深度安全阈值

写在便利贴，下次架构评审你即Agent架构专家！