OpenClaw：本地 AI 代理的架构、开发与安全
本课程深入分析了 OpenClaw——一个开创性的开源框架，专为自主 AI 代理设计。课程系统性地剖析了该框架的分层系统架构、本地优先的 RAG 内存机制、浏览器自动化协议以及高度可扩展的技能生态系统。教学内容涵盖复杂工作流的实际编排，包括 PIV 自动化流程和多智能体委员会模式。此外，课程还深入分析了在生产级部署范式中的硬件权衡，并提出了针对核心安全威胁（如远程代码执行漏洞和提示注入）的纵深防御策略。本课程旨在赋能高级开发人员与架构师，构建具备高自主性的同时仍保持安全可控的 AI 代理系统。
课程概述
📚 内容概要
本课程深入剖析OpenClaw——一个开创性的开源框架，专为自主AI代理设计。我们系统性地解构其分层架构，重点聚焦于本地优先的RAG记忆机制、浏览器自动化协议以及高度可扩展的技能生态系统。
课程内容超越理论，涵盖复杂工作流的实际编排，例如PIV自动化流程和多智能体委员会模式。尤为重要的是，课程针对生产环境中的关键挑战展开分析，权衡硬件取舍，并实施纵深防御策略，以应对诸如远程代码执行（RCE）漏洞和提示注入等严重威胁。本课程旨在赋能资深开发者与系统架构师，构建既高度自治又严格安全的AI系统。
目标受众：资深开发者与系统架构师
🎯 学习目标
完成本课程后，您将能够：
- 设计基于OpenClaw框架及其“Markdown优先”理念的自主系统。
- 部署安全的本地优先内存架构，防止高并发环境下的状态损坏。
- 强化代理供应链，抵御包括间接提示注入和静默回退型RCE在内的高级威胁。
课程 共 5 课时 · 预计 15.0h
课程
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.