AI Agent 怎么学？保姆级 12 步学习路线来了（附 15 篇权威论文）AI Agent 怎么学？保姆级 12 步学习路线来了（附 15 篇权威论文）
很多人觉得 AI Agent 门槛高、资料又碎又杂。市面上的内容要么停在"概念科普"，要么一上来就堆框架，新手很难串成一条能落地的路。为此，AI 导航整理了一条「从零基础到能落地」的清晰学习路线，把 Agent 拆成 12 个可执行的步骤。
路线总览：第 1–2 步建立概念与核心思维（规划 + 记忆）；第 3–5 步是动手三件套——Prompt 工程、RAG 检索增强、工具调用 / Function Calling / MCP；第 6–7 步讲设计模式（ReAct / Plan-and-Solve / Reflexion）与多 Agent 协作；第 8–9 步是评估与安全（AgentBench、LLM-as-a-Judge、OWASP LLM Top10）；第 10–12 步落到部署、实战与持续学习。开篇还附一张「思考-行动」循环总览图，先建立全局画面。
权威加持是这次教程的重点。每一步都标注了奠基论文与官方规范：RAG（Lewis 2020）、ReAct（Yao 2022）、Chain-of-Thought（Wei 2022）、Tree of Thoughts（Yao 2023）、Reflexion（Shinn 2023）、Generative Agents（Park 2023）、AgentBench（Liu 2023）、LLM-as-a-Judge（Zheng 2023），以及 MCP、OpenAI、Anthropic、LangChain、AutoGen、Lilian Weng 综述、OWASP 等可核验站点。所有引用只给标题 / 作者 / 年份与官方域名，绝不编造链接。
光有理论不够，教程是「保姆级」的：每步配可复制的 Python 代码（OpenAI 兼容接口，DeepSeek / 硅基流动等都能跑）、预期输出、常见坑提醒和一张自测清单。例如第 4 步 RAG 直接给出可离线运行的本地 BGE 向量化方案（中文友好、无需 Key）；第 10 步给出完整 FastAPI 封装 + Dockerfile + requirements.txt，照抄即可部署。
教程完全免费阅读，文末还附了 5 篇站内互链（企微 × dsh 实战、dsh 基础、WorkBuddy、什么是 CLI、GitHub 教程），与「企业微信全面开放 CLI/MCP」那条资讯形成系列。无论你是想入门、还是想把零散知识系统串起来，这条 12 步路线都值得收藏。
Many people find AI Agents intimidating — the material is either too high-level or jumps straight into frameworks. AI Navigator has published a structured 12-step learning path from zero to a working Agent.
Overview: Steps 1–2 build concepts and core thinking (planning + memory); 3–5 cover Prompt engineering, RAG, and tool use / Function Calling / MCP; 6–7 cover design patterns (ReAct / Plan-and-Solve / Reflexion) and multi-agent; 8–9 cover evaluation (AgentBench, LLM-as-a-Judge) and security (OWASP LLM Top 10); 10–12 cover deployment, real projects, and continuous learning. An "think-act loop" diagram opens the guide.
Every step cites foundational papers and official specs: RAG (Lewis 2020), ReAct (Yao 2022), Chain-of-Thought (Wei 2022), Tree of Thoughts (Yao 2023), Reflexion (Shinn 2023), Generative Agents (Park 2023), AgentBench (Liu 2023), LLM-as-a-Judge (Zheng 2023), plus verifiable sites like modelcontextprotocol.io, platform.openai.com/docs, docs.anthropic.com, docs.langchain.com, github.com/microsoft/autogen, lilianweng.github.io, owasp.org.
It's hands-on: each step ships copy-pasteable Python (OpenAI-compatible, runs on DeepSeek / SiliconFlow), expected output, pitfalls, and a self-check list. Step 4 gives an offline BGE embedding setup (Chinese-friendly, no API key); Step 10 ships a full FastAPI + Dockerfile + requirements.txt.
Free to read, with 5 cross-linked tutorials at the end. A good one to bookmark whether you're starting out or consolidating.
打开微信「扫一扫」，分享给好友或朋友圈
关闭