从0到1开发一个AI Agent：完整实战指南

/* MD / 富文本中的 .toc（含博客园搬家等嵌套结构）；.toc-box 在侧栏，不受影响 */
#content_views .toc,
/* 编辑器常在目录前后插入空 p（:empty 仍占 20px），一并去掉避免顶空隙 */
#content_views.markdown_views > p:empty:has(+ .toc),
#content_views.markdown_views > .toc + p:empty,
/* 富文本旧版目录标记 */
#content_views.htmledit_views #main-toc,
#content_views.htmledit_views #hr-toc,
#content_views.htmledit_views p[id*=\”-toc\”] {
display: none !important;
}
/* 目录去掉后，紧跟的首个标题不再多出一块上边距 */
#content_views.markdown_views > .toc + h1,
#content_views.markdown_views > .toc + h2,
#content_views.markdown_views > .toc + h3,
#content_views.markdown_views > .toc + h4,
#content_views.markdown_views > .toc + p:empty + h1,
#content_views.markdown_views > .toc + p:empty + h2,
#content_views.markdown_views > .toc + p:empty + h3,
#content_views.markdown_views > .toc + p:empty + h4 {
margin-top: 0 !important;
}

1. 引言

下面这张图展示了 AI Agent 的整体架构，帮助你快速建立全局认知：

#mermaid-svg-yFI9AujoCoKnw1N7{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-yFI9AujoCoKnw1N7 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-yFI9AujoCoKnw1N7 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-yFI9AujoCoKnw1N7 .error-icon{fill:#552222;}#mermaid-svg-yFI9AujoCoKnw1N7 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-yFI9AujoCoKnw1N7 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-yFI9AujoCoKnw1N7 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-yFI9AujoCoKnw1N7 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-yFI9AujoCoKnw1N7 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-yFI9AujoCoKnw1N7 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-yFI9AujoCoKnw1N7 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-yFI9AujoCoKnw1N7 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-yFI9AujoCoKnw1N7 .marker.cross{stroke:#333333;}#mermaid-svg-yFI9AujoCoKnw1N7 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-yFI9AujoCoKnw1N7 p{margin:0;}#mermaid-svg-yFI9AujoCoKnw1N7 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-yFI9AujoCoKnw1N7 .cluster-label text{fill:#333;}#mermaid-svg-yFI9AujoCoKnw1N7 .cluster-label span{color:#333;}#mermaid-svg-yFI9AujoCoKnw1N7 .cluster-label span p{background-color:transparent;}#mermaid-svg-yFI9AujoCoKnw1N7 .label text,#mermaid-svg-yFI9AujoCoKnw1N7 span{fill:#333;color:#333;}#mermaid-svg-yFI9AujoCoKnw1N7 .node rect,#mermaid-svg-yFI9AujoCoKnw1N7 .node circle,#mermaid-svg-yFI9AujoCoKnw1N7 .node ellipse,#mermaid-svg-yFI9AujoCoKnw1N7 .node polygon,#mermaid-svg-yFI9AujoCoKnw1N7 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-yFI9AujoCoKnw1N7 .rough-node .label text,#mermaid-svg-yFI9AujoCoKnw1N7 .node .label text,#mermaid-svg-yFI9AujoCoKnw1N7 .image-shape .label,#mermaid-svg-yFI9AujoCoKnw1N7 .icon-shape .label{text-anchor:middle;}#mermaid-svg-yFI9AujoCoKnw1N7 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-yFI9AujoCoKnw1N7 .rough-node .label,#mermaid-svg-yFI9AujoCoKnw1N7 .node .label,#mermaid-svg-yFI9AujoCoKnw1N7 .image-shape .label,#mermaid-svg-yFI9AujoCoKnw1N7 .icon-shape .label{text-align:center;}#mermaid-svg-yFI9AujoCoKnw1N7 .node.clickable{cursor:pointer;}#mermaid-svg-yFI9AujoCoKnw1N7 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-yFI9AujoCoKnw1N7 .arrowheadPath{fill:#333333;}#mermaid-svg-yFI9AujoCoKnw1N7 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-yFI9AujoCoKnw1N7 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-yFI9AujoCoKnw1N7 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-yFI9AujoCoKnw1N7 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-yFI9AujoCoKnw1N7 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-yFI9AujoCoKnw1N7 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-yFI9AujoCoKnw1N7 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-yFI9AujoCoKnw1N7 .cluster text{fill:#333;}#mermaid-svg-yFI9AujoCoKnw1N7 .cluster span{color:#333;}#mermaid-svg-yFI9AujoCoKnw1N7 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-yFI9AujoCoKnw1N7 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-yFI9AujoCoKnw1N7 rect.text{fill:none;stroke-width:0;}#mermaid-svg-yFI9AujoCoKnw1N7 .icon-shape,#mermaid-svg-yFI9AujoCoKnw1N7 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-yFI9AujoCoKnw1N7 .icon-shape p,#mermaid-svg-yFI9AujoCoKnw1N7 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-yFI9AujoCoKnw1N7 .icon-shape .label rect,#mermaid-svg-yFI9AujoCoKnw1N7 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-yFI9AujoCoKnw1N7 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-yFI9AujoCoKnw1N7 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-yFI9AujoCoKnw1N7 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

需要外部信息

可以回答

👤 用户输入

🧠 大语言模型（LLM）

🤔 思考与规划

🔧 工具调用

🌐 网络搜索

💾 数据库查询

🧮 代码执行器

📁 文件操作

📥 工具结果

💬 生成最终答案

📝 记忆存储

AI Agent（智能体）是当前人工智能领域最热门的概念之一。它不再是一个简单的问答模型，而是一个能够自主理解目标、拆解任务、调用工具、并执行多步操作的智能系统。本文将带你从零开始，手把手构建一个属于自己的AI Agent，让你深入理解其核心原理与实现方式。

2. 什么是AI Agent？

下面这张图展示了 AI Agent 的五大核心能力及其协作关系：

#mermaid-svg-ytXmJuqYC5z5jUya{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-ytXmJuqYC5z5jUya .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-ytXmJuqYC5z5jUya .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-ytXmJuqYC5z5jUya .error-icon{fill:#552222;}#mermaid-svg-ytXmJuqYC5z5jUya .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-ytXmJuqYC5z5jUya .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-ytXmJuqYC5z5jUya .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-ytXmJuqYC5z5jUya .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-ytXmJuqYC5z5jUya .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-ytXmJuqYC5z5jUya .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-ytXmJuqYC5z5jUya .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-ytXmJuqYC5z5jUya .marker{fill:#333333;stroke:#333333;}#mermaid-svg-ytXmJuqYC5z5jUya .marker.cross{stroke:#333333;}#mermaid-svg-ytXmJuqYC5z5jUya svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-ytXmJuqYC5z5jUya p{margin:0;}#mermaid-svg-ytXmJuqYC5z5jUya .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-ytXmJuqYC5z5jUya .cluster-label text{fill:#333;}#mermaid-svg-ytXmJuqYC5z5jUya .cluster-label span{color:#333;}#mermaid-svg-ytXmJuqYC5z5jUya .cluster-label span p{background-color:transparent;}#mermaid-svg-ytXmJuqYC5z5jUya .label text,#mermaid-svg-ytXmJuqYC5z5jUya span{fill:#333;color:#333;}#mermaid-svg-ytXmJuqYC5z5jUya .node rect,#mermaid-svg-ytXmJuqYC5z5jUya .node circle,#mermaid-svg-ytXmJuqYC5z5jUya .node ellipse,#mermaid-svg-ytXmJuqYC5z5jUya .node polygon,#mermaid-svg-ytXmJuqYC5z5jUya .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-ytXmJuqYC5z5jUya .rough-node .label text,#mermaid-svg-ytXmJuqYC5z5jUya .node .label text,#mermaid-svg-ytXmJuqYC5z5jUya .image-shape .label,#mermaid-svg-ytXmJuqYC5z5jUya .icon-shape .label{text-anchor:middle;}#mermaid-svg-ytXmJuqYC5z5jUya .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-ytXmJuqYC5z5jUya .rough-node .label,#mermaid-svg-ytXmJuqYC5z5jUya .node .label,#mermaid-svg-ytXmJuqYC5z5jUya .image-shape .label,#mermaid-svg-ytXmJuqYC5z5jUya .icon-shape .label{text-align:center;}#mermaid-svg-ytXmJuqYC5z5jUya .node.clickable{cursor:pointer;}#mermaid-svg-ytXmJuqYC5z5jUya .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-ytXmJuqYC5z5jUya .arrowheadPath{fill:#333333;}#mermaid-svg-ytXmJuqYC5z5jUya .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-ytXmJuqYC5z5jUya .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-ytXmJuqYC5z5jUya .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-ytXmJuqYC5z5jUya .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-ytXmJuqYC5z5jUya .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-ytXmJuqYC5z5jUya .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-ytXmJuqYC5z5jUya .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-ytXmJuqYC5z5jUya .cluster text{fill:#333;}#mermaid-svg-ytXmJuqYC5z5jUya .cluster span{color:#333;}#mermaid-svg-ytXmJuqYC5z5jUya div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-ytXmJuqYC5z5jUya .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-ytXmJuqYC5z5jUya rect.text{fill:none;stroke-width:0;}#mermaid-svg-ytXmJuqYC5z5jUya .icon-shape,#mermaid-svg-ytXmJuqYC5z5jUya .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-ytXmJuqYC5z5jUya .icon-shape p,#mermaid-svg-ytXmJuqYC5z5jUya .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-ytXmJuqYC5z5jUya .icon-shape .label rect,#mermaid-svg-ytXmJuqYC5z5jUya .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-ytXmJuqYC5z5jUya .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-ytXmJuqYC5z5jUya .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-ytXmJuqYC5z5jUya :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

🤖 AI Agent 核心能力

反馈循环

👁️ 感知Perception

🧠 规划Planning

💾 记忆Memory

🔧 工具使用Tool Use

⚡ 行动Action

🌍 外部环境

👤 用户

2.1 与传统问答系统的区别

在开始编码之前，我们先明确AI Agent的定义。一个典型的AI Agent具备以下核心能力：

• 感知（Perception）：接收并理解用户输入的指令或环境信息。
• 规划（Planning）：将复杂目标拆解为一系列可执行的子任务。
• 记忆（Memory）：存储对话历史、中间结果和关键信息。
• 工具使用（Tool Use）：调用外部API、数据库、搜索引擎或代码执行器来完成任务。
• 行动（Action）：执行具体的操作，并基于反馈调整后续行为。

简单来说，AI Agent = 大语言模型（LLM） + 规划能力 + 记忆系统 + 工具集。

3. 环境准备与项目初始化

在开始编码之前，我们先搭建好完整的开发环境。下面是整个项目的目录结构和依赖关系图：

#mermaid-svg-4sCyeB3KFDzexP6F{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-4sCyeB3KFDzexP6F .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-4sCyeB3KFDzexP6F .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-4sCyeB3KFDzexP6F .error-icon{fill:#552222;}#mermaid-svg-4sCyeB3KFDzexP6F .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-4sCyeB3KFDzexP6F .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-4sCyeB3KFDzexP6F .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-4sCyeB3KFDzexP6F .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-4sCyeB3KFDzexP6F .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-4sCyeB3KFDzexP6F .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-4sCyeB3KFDzexP6F .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-4sCyeB3KFDzexP6F .marker{fill:#333333;stroke:#333333;}#mermaid-svg-4sCyeB3KFDzexP6F .marker.cross{stroke:#333333;}#mermaid-svg-4sCyeB3KFDzexP6F svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-4sCyeB3KFDzexP6F p{margin:0;}#mermaid-svg-4sCyeB3KFDzexP6F .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-4sCyeB3KFDzexP6F .cluster-label text{fill:#333;}#mermaid-svg-4sCyeB3KFDzexP6F .cluster-label span{color:#333;}#mermaid-svg-4sCyeB3KFDzexP6F .cluster-label span p{background-color:transparent;}#mermaid-svg-4sCyeB3KFDzexP6F .label text,#mermaid-svg-4sCyeB3KFDzexP6F span{fill:#333;color:#333;}#mermaid-svg-4sCyeB3KFDzexP6F .node rect,#mermaid-svg-4sCyeB3KFDzexP6F .node circle,#mermaid-svg-4sCyeB3KFDzexP6F .node ellipse,#mermaid-svg-4sCyeB3KFDzexP6F .node polygon,#mermaid-svg-4sCyeB3KFDzexP6F .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-4sCyeB3KFDzexP6F .rough-node .label text,#mermaid-svg-4sCyeB3KFDzexP6F .node .label text,#mermaid-svg-4sCyeB3KFDzexP6F .image-shape .label,#mermaid-svg-4sCyeB3KFDzexP6F .icon-shape .label{text-anchor:middle;}#mermaid-svg-4sCyeB3KFDzexP6F .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-4sCyeB3KFDzexP6F .rough-node .label,#mermaid-svg-4sCyeB3KFDzexP6F .node .label,#mermaid-svg-4sCyeB3KFDzexP6F .image-shape .label,#mermaid-svg-4sCyeB3KFDzexP6F .icon-shape .label{text-align:center;}#mermaid-svg-4sCyeB3KFDzexP6F .node.clickable{cursor:pointer;}#mermaid-svg-4sCyeB3KFDzexP6F .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-4sCyeB3KFDzexP6F .arrowheadPath{fill:#333333;}#mermaid-svg-4sCyeB3KFDzexP6F .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-4sCyeB3KFDzexP6F .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-4sCyeB3KFDzexP6F .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-4sCyeB3KFDzexP6F .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-4sCyeB3KFDzexP6F .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-4sCyeB3KFDzexP6F .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-4sCyeB3KFDzexP6F .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-4sCyeB3KFDzexP6F .cluster text{fill:#333;}#mermaid-svg-4sCyeB3KFDzexP6F .cluster span{color:#333;}#mermaid-svg-4sCyeB3KFDzexP6F div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-4sCyeB3KFDzexP6F .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-4sCyeB3KFDzexP6F rect.text{fill:none;stroke-width:0;}#mermaid-svg-4sCyeB3KFDzexP6F .icon-shape,#mermaid-svg-4sCyeB3KFDzexP6F .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-4sCyeB3KFDzexP6F .icon-shape p,#mermaid-svg-4sCyeB3KFDzexP6F .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-4sCyeB3KFDzexP6F .icon-shape .label rect,#mermaid-svg-4sCyeB3KFDzexP6F .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-4sCyeB3KFDzexP6F .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-4sCyeB3KFDzexP6F .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-4sCyeB3KFDzexP6F :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

📁 ai-agent-tutorial/

main.py主程序入口

agent_core.pyAgent核心逻辑

tools.py工具定义

memory.py记忆管理

.env环境变量

requirements.txt依赖清单

🐍 Python 3.10+

🔑 OpenAI API

🌐 DuckDuckGo

3.1 技术选型

我们将使用以下技术栈：

• Python 3.10+：开发语言
• OpenAI API：作为核心LLM（兼容任何支持Function Calling的模型）
• LangChain：作为Agent框架（可选，但能极大简化开发）
• DuckDuckGo Search：作为示例工具

3.2 安装依赖

mkdir ai-agent-tutorial
cd ai-agent-tutorial
python -m venv venv
source venv/bin/activate # Windows: venv\\Scripts\\activate

pip install openai langchain langchain-community duckduckgo-search python-dotenv

3.3 配置环境变量

在项目根目录创建 .env 文件：

OPENAI_API_KEY=your_openai_api_key_here

4. 核心实现：从零构建Agent

下面这张图展示了我们将要实现的三个核心模块及其演进关系：

#mermaid-svg-9RwM8pn2HdCnMnWb{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-9RwM8pn2HdCnMnWb .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-9RwM8pn2HdCnMnWb .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-9RwM8pn2HdCnMnWb .error-icon{fill:#552222;}#mermaid-svg-9RwM8pn2HdCnMnWb .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-9RwM8pn2HdCnMnWb .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-9RwM8pn2HdCnMnWb .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-9RwM8pn2HdCnMnWb .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-9RwM8pn2HdCnMnWb .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-9RwM8pn2HdCnMnWb .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-9RwM8pn2HdCnMnWb .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-9RwM8pn2HdCnMnWb .marker{fill:#333333;stroke:#333333;}#mermaid-svg-9RwM8pn2HdCnMnWb .marker.cross{stroke:#333333;}#mermaid-svg-9RwM8pn2HdCnMnWb svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-9RwM8pn2HdCnMnWb p{margin:0;}#mermaid-svg-9RwM8pn2HdCnMnWb .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-9RwM8pn2HdCnMnWb .cluster-label text{fill:#333;}#mermaid-svg-9RwM8pn2HdCnMnWb .cluster-label span{color:#333;}#mermaid-svg-9RwM8pn2HdCnMnWb .cluster-label span p{background-color:transparent;}#mermaid-svg-9RwM8pn2HdCnMnWb .label text,#mermaid-svg-9RwM8pn2HdCnMnWb span{fill:#333;color:#333;}#mermaid-svg-9RwM8pn2HdCnMnWb .node rect,#mermaid-svg-9RwM8pn2HdCnMnWb .node circle,#mermaid-svg-9RwM8pn2HdCnMnWb .node ellipse,#mermaid-svg-9RwM8pn2HdCnMnWb .node polygon,#mermaid-svg-9RwM8pn2HdCnMnWb .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-9RwM8pn2HdCnMnWb .rough-node .label text,#mermaid-svg-9RwM8pn2HdCnMnWb .node .label text,#mermaid-svg-9RwM8pn2HdCnMnWb .image-shape .label,#mermaid-svg-9RwM8pn2HdCnMnWb .icon-shape .label{text-anchor:middle;}#mermaid-svg-9RwM8pn2HdCnMnWb .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-9RwM8pn2HdCnMnWb .rough-node .label,#mermaid-svg-9RwM8pn2HdCnMnWb .node .label,#mermaid-svg-9RwM8pn2HdCnMnWb .image-shape .label,#mermaid-svg-9RwM8pn2HdCnMnWb .icon-shape .label{text-align:center;}#mermaid-svg-9RwM8pn2HdCnMnWb .node.clickable{cursor:pointer;}#mermaid-svg-9RwM8pn2HdCnMnWb .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-9RwM8pn2HdCnMnWb .arrowheadPath{fill:#333333;}#mermaid-svg-9RwM8pn2HdCnMnWb .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-9RwM8pn2HdCnMnWb .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-9RwM8pn2HdCnMnWb .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-9RwM8pn2HdCnMnWb .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-9RwM8pn2HdCnMnWb .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-9RwM8pn2HdCnMnWb .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-9RwM8pn2HdCnMnWb .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-9RwM8pn2HdCnMnWb .cluster text{fill:#333;}#mermaid-svg-9RwM8pn2HdCnMnWb .cluster span{color:#333;}#mermaid-svg-9RwM8pn2HdCnMnWb div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-9RwM8pn2HdCnMnWb .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-9RwM8pn2HdCnMnWb rect.text{fill:none;stroke-width:0;}#mermaid-svg-9RwM8pn2HdCnMnWb .icon-shape,#mermaid-svg-9RwM8pn2HdCnMnWb .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-9RwM8pn2HdCnMnWb .icon-shape p,#mermaid-svg-9RwM8pn2HdCnMnWb .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-9RwM8pn2HdCnMnWb .icon-shape .label rect,#mermaid-svg-9RwM8pn2HdCnMnWb .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-9RwM8pn2HdCnMnWb .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-9RwM8pn2HdCnMnWb .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-9RwM8pn2HdCnMnWb :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

步骤三：加入记忆系统

短期记忆

长期记忆

上下文管理

步骤二：引入工具调用

Function Calling

搜索工具

计算工具

其他工具

步骤一：基础ReAct循环

思考（Reasoning）

行动（Acting）

我们将分三步实现：先实现一个基础的ReAct循环，再引入工具调用，最后加入记忆功能。

4.1 基础ReAct循环

ReAct 循环是 Agent 最核心的执行范式。下面这张图展示了它的完整执行流程：

#mermaid-svg-5v0xFc4wUkFezR9n{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-5v0xFc4wUkFezR9n .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-5v0xFc4wUkFezR9n .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-5v0xFc4wUkFezR9n .error-icon{fill:#552222;}#mermaid-svg-5v0xFc4wUkFezR9n .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-5v0xFc4wUkFezR9n .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-5v0xFc4wUkFezR9n .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-5v0xFc4wUkFezR9n .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-5v0xFc4wUkFezR9n .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-5v0xFc4wUkFezR9n .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-5v0xFc4wUkFezR9n .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-5v0xFc4wUkFezR9n .marker{fill:#333333;stroke:#333333;}#mermaid-svg-5v0xFc4wUkFezR9n .marker.cross{stroke:#333333;}#mermaid-svg-5v0xFc4wUkFezR9n svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-5v0xFc4wUkFezR9n p{margin:0;}#mermaid-svg-5v0xFc4wUkFezR9n .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-5v0xFc4wUkFezR9n .cluster-label text{fill:#333;}#mermaid-svg-5v0xFc4wUkFezR9n .cluster-label span{color:#333;}#mermaid-svg-5v0xFc4wUkFezR9n .cluster-label span p{background-color:transparent;}#mermaid-svg-5v0xFc4wUkFezR9n .label text,#mermaid-svg-5v0xFc4wUkFezR9n span{fill:#333;color:#333;}#mermaid-svg-5v0xFc4wUkFezR9n .node rect,#mermaid-svg-5v0xFc4wUkFezR9n .node circle,#mermaid-svg-5v0xFc4wUkFezR9n .node ellipse,#mermaid-svg-5v0xFc4wUkFezR9n .node polygon,#mermaid-svg-5v0xFc4wUkFezR9n .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-5v0xFc4wUkFezR9n .rough-node .label text,#mermaid-svg-5v0xFc4wUkFezR9n .node .label text,#mermaid-svg-5v0xFc4wUkFezR9n .image-shape .label,#mermaid-svg-5v0xFc4wUkFezR9n .icon-shape .label{text-anchor:middle;}#mermaid-svg-5v0xFc4wUkFezR9n .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-5v0xFc4wUkFezR9n .rough-node .label,#mermaid-svg-5v0xFc4wUkFezR9n .node .label,#mermaid-svg-5v0xFc4wUkFezR9n .image-shape .label,#mermaid-svg-5v0xFc4wUkFezR9n .icon-shape .label{text-align:center;}#mermaid-svg-5v0xFc4wUkFezR9n .node.clickable{cursor:pointer;}#mermaid-svg-5v0xFc4wUkFezR9n .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-5v0xFc4wUkFezR9n .arrowheadPath{fill:#333333;}#mermaid-svg-5v0xFc4wUkFezR9n .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-5v0xFc4wUkFezR9n .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-5v0xFc4wUkFezR9n .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-5v0xFc4wUkFezR9n .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-5v0xFc4wUkFezR9n .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-5v0xFc4wUkFezR9n .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-5v0xFc4wUkFezR9n .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-5v0xFc4wUkFezR9n .cluster text{fill:#333;}#mermaid-svg-5v0xFc4wUkFezR9n .cluster span{color:#333;}#mermaid-svg-5v0xFc4wUkFezR9n div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-5v0xFc4wUkFezR9n .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-5v0xFc4wUkFezR9n rect.text{fill:none;stroke-width:0;}#mermaid-svg-5v0xFc4wUkFezR9n .icon-shape,#mermaid-svg-5v0xFc4wUkFezR9n .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-5v0xFc4wUkFezR9n .icon-shape p,#mermaid-svg-5v0xFc4wUkFezR9n .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-5v0xFc4wUkFezR9n .icon-shape .label rect,#mermaid-svg-5v0xFc4wUkFezR9n .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-5v0xFc4wUkFezR9n .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-5v0xFc4wUkFezR9n .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-5v0xFc4wUkFezR9n :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

不能，需要行动

能

🚀 开始

📝 接收用户输入

🤔 思考：分析问题确定需要什么信息

🧐 判断：能否直接回答？

⚡ 行动：执行某个操作

👀 观察：获取行动结果

💬 生成最终答案

✅ 结束

ReAct 循环的关键特点：

ReAct（Reasoning + Acting）是Agent最核心的范式。它让LLM在“思考”和“行动”之间交替，直到得出最终答案。

import os
from openai import OpenAI
from dotenv import load_dotenv

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

def simple_agent(user_input: str, max_steps: int = 5):
messages = [
{"role": "system", "content": "你是一个AI助手。请逐步思考并行动，最终给出答案。"},
{"role": "user", "content": user_input}
]

for step in range(max_steps):
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=messages
)
assistant_message = response.choices[0].message
messages.append(assistant_message)

# 检查是否已给出最终答案
if assistant_message.content and "最终答案" in assistant_message.content:
return assistant_message.content

# 模拟思考-行动循环
print(f"Step {step + 1}: {assistant_message.content}")

return "已达到最大步数，未得出最终答案。"

# 测试
result = simple_agent("请计算 25 的平方根，并告诉我结果。")
print(result)

4.2 引入工具调用（Function Calling）

真正的Agent需要能调用外部工具。OpenAI的Function Calling机制让LLM可以请求执行特定函数。

首先，定义一个搜索工具：

import json
from duckduckgo_search import DDGS

def search_web(query: str) –> str:
"""使用DuckDuckGo搜索网络信息"""
with DDGS() as ddgs:
results = list(ddgs.text(query, max_results=3))
if results:
return "\\n".join([f"{r['title']}: {r['body']}" for r in results])
return "未找到相关结果。"

# 定义工具描述（用于Function Calling）
tools = [
{
"type": "function",
"function": {
"name": "search_web",
"description": "搜索网络获取最新信息",
"parameters": {
"type": "object",
"properties": {
"query": {
"type": "string",
"description": "搜索关键词"
}
},
"required": ["query"]
}
}
}
]

# 工具映射表
tool_map = {
"search_web": search_web
}

然后，实现带工具调用的Agent循环：

def agent_with_tools(user_input: str, max_steps: int = 5):
messages = [
{"role": "system", "content": "你是一个能调用工具的AI助手。请根据用户需求，合理调用工具并给出最终答案。"},
{"role": "user", "content": user_input}
]

for step in range(max_steps):
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=messages,
tools=tools,
tool_choice="auto"
)

assistant_message = response.choices[0].message
messages.append(assistant_message)

# 检查是否有工具调用请求
if assistant_message.tool_calls:
for tool_call in assistant_message.tool_calls:
function_name = tool_call.function.name
function_args = json.loads(tool_call.function.arguments)

print(f"🔧 调用工具: {function_name}({function_args})")

# 执行工具
function_response = tool_map[function_name](**function_args)

# 将工具结果返回给LLM
messages.append({
"role": "tool",
"tool_call_id": tool_call.id,
"content": function_response
})
else:
# 没有工具调用，说明LLM给出了最终答案
return assistant_message.content

return "已达到最大步数。"

# 测试
result = agent_with_tools("2024年巴黎奥运会中国获得了多少枚金牌？")
print(f"最终答案: {result}")

4.3 加入记忆系统

为了让Agent能记住对话上下文，我们需要实现一个简单的记忆管理：

from typing import List, Dict

class SimpleMemory:
def __init__(self, max_history: int = 10):
self.history: List[Dict] = []
self.max_history = max_history

def add_message(self, role: str, content: str):
self.history.append({"role": role, "content": content})
# 限制历史长度
if len(self.history) > self.max_history:
self.history = self.history[–self.max_history:]

def get_context(self) –> List[Dict]:
return self.history

def clear(self):
self.history = []

def agent_with_memory(user_input: str, memory: SimpleMemory, max_steps: int = 5):
# 构建消息列表：系统提示 + 历史记忆 + 当前输入
messages = [
{"role": "system", "content": "你是一个有记忆的AI助手。请记住对话上下文，合理调用工具。"}
]
messages.extend(memory.get_context())
messages.append({"role": "user", "content": user_input})

for step in range(max_steps):
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=messages,
tools=tools,
tool_choice="auto"
)

assistant_message = response.choices[0].message
messages.append(assistant_message)

if assistant_message.tool_calls:
for tool_call in assistant_message.tool_calls:
function_name = tool_call.function.name
function_args = json.loads(tool_call.function.arguments)
function_response = tool_map[function_name](**function_args)

messages.append({
"role": "tool",
"tool_call_id": tool_call.id,
"content": function_response
})
else:
# 保存到记忆
memory.add_message("user", user_input)
memory.add_message("assistant", assistant_message.content)
return assistant_message.content

return "已达到最大步数。"

# 测试多轮对话
memory = SimpleMemory()
print(agent_with_memory("你好，我叫小明", memory))
print(agent_with_memory("我刚才说了什么？", memory)) # 应该记得名字

5. 进阶：使用LangChain快速构建

如果你不想从零实现，LangChain提供了开箱即用的Agent构建能力：

from langchain.agents import create_react_agent, AgentExecutor
from langchain.tools import Tool
from langchain_openai import ChatOpenAI
from langchain.prompts import PromptTemplate

# 定义工具
def search_tool(query: str) –> str:
with DDGS() as ddgs:
results = list(ddgs.text(query, max_results=3))
return "\\n".join([r["body"] for r in results]) if results else "无结果"

tools = [
Tool(
name="web_search",
func=search_tool,
description="搜索网络信息，输入为搜索关键词"
)
]

# 创建Agent
llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
prompt = PromptTemplate.from_template(
"你是一个AI助手。请使用以下工具回答问题：\\n{tools}\\n\\n"
"工具名称: {tool_names}\\n"
"用户问题: {input}\\n\\n"
"请逐步思考：\\n{agent_scratchpad}"
)

agent = create_react_agent(llm, tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)

# 运行
result = agent_executor.invoke({"input": "今天比特币价格是多少？"})
print(result["output"])

6. 部署与优化建议

6.1 生产环境注意事项

• 错误处理：为每个工具调用添加try-except，防止单点故障导致整个Agent崩溃。
• 超时控制：为LLM调用和工具执行设置超时时间。
• 速率限制：实现令牌桶或滑动窗口算法，防止API调用超限。
• 日志记录：记录每次思考-行动循环，便于调试和审计。

6.2 性能优化

• 缓存：对相同输入的LLM响应和工具结果进行缓存。
• 并行工具调用：对于不依赖顺序的工具，可以并行执行。
• 精简上下文：使用滑动窗口或摘要压缩历史对话，避免超出上下文窗口。

7. 总结

本文从零开始，带你完整实现了一个具备工具调用和记忆能力的AI Agent。核心要点总结如下：

现在，你已经掌握了构建AI Agent的核心技能。尝试为你的Agent添加更多工具（如计算器、数据库查询、文件操作），让它变得更加强大吧！