OpenClaw：本地 AI 代理的架構、開發與安全
本課程深入分析 OpenClaw，這是一個革命性的開源框架，專為自主型人工智慧代理設計。課程系統性地拆解該框架的層疊式架構、以本地為首的 RAG 記憶機制、瀏覽器自動化協定以及高度可擴展的技能生態系。課程內容涵蓋複雜工作流程的實際整合，包括 PIV 自動化流程與多代理委員會模式。此外，還批判性地分析生產級部署模式中的硬體取捨，並提出針對核心安全威脅（如遠端程式碼執行漏洞與提示注入）的纵深防禦策略。本課程旨在賦能資深開發者與架構師，建立具有高度自主性，同時仍具備安全性與可控性的 AI 代理系統。
課程總覽
📚 內容摘要
本課程深入探討 OpenClaw——一個革命性的開源框架，專為自主式人工智慧代理設計。我們系統性地剖析其分層架構，重點聚焦於 本地優先的 RAG 記憶機制、瀏覽器自動化協定，以及高度可擴展的技能生態系統。
課程內容不僅涵蓋理論，更著重於複雜工作流程的實際整合，例如 PIV 自動化流程 和 多代理委員會模式。關鍵的是，課程解決了生產環境中的挑戰，分析硬體取捨並實施 深度防禦策略，以應對如遠端程式碼執行（RCE）漏洞和提示注入等重大威脅。本課程旨在賦能資深開發者與系統架構師，打造既高度自主又嚴謹安全的人工智慧系統。
目標受眾：資深開發者與系統架構師
🎯 學習目標
完成本課程後，您將能夠：
- 架構 使用 OpenClaw 框架及其「Markdown 為先」哲學的自主系統。
- 部署 安全、本地優先的記憶體架構，在高併發環境中防止狀態損壞。
- 加固 代理供應鏈，抵禦如間接提示注入與靜默回退 RCE 等進階威脅。
課程 共 5 课时 · 预计 15.0h
課程
Lesson
This lesson introduces the OpenClaw architecture, which utilizes a Markdown-first philosophy to define agent identity and logic through human-readable files like SOUL.md and MEMORY.md. Students learn how the system ensures reliability and security through session isolation via lane queues, a model-agnostic runtime, and a centralized gateway for managing autonomous tasks.
This lesson explores the OpenClaw architecture, focusing on the PIV (Planning-Interaction-Verification) workflow for autonomous browser agents and the importance of session isolation for system security. It also covers the modular configuration system, where openclaw.json manages technical routing and SOUL.md defines the agent's persistent identity and ethical boundaries.
This lesson introduces the OpenClaw architecture, which utilizes a Markdown-first philosophy to manage agent identity, capabilities, and workflows through human-readable configuration files. Students learn how to build robust, model-agnostic agents by leveraging core infrastructure components like the Gateway, Agent Runtime, and a structured manifest system to ensure secure, version-controlled, and transparent AI orchestration.
This lesson covers the production deployment of autonomous agents using the OpenClaw architecture, emphasizing the Markdown-First philosophy for configuration and runtime sovereignty. Students will learn to optimize performance through hybrid execution models, local-first RAG, and memory management techniques like semantic snapshots and pre-compaction flushing.
This lesson explores the evolving threat landscape for autonomous AI agents, focusing on risks like indirect prompt injection, skill supply chain poisoning, and delegated compromise. It emphasizes implementing a Zero Trust architecture, including strict network binding and ephemeral credential management, to secure agent gateways and minimize the blast radius of potential security breaches.