Open Claw 的创造者 Peter Steinberger 讨论构建一个席卷互联网的自主 AI 助手，他从 PSPDFKit 到代理工程的历程，以及 AI 代理的未来。
Peter Steinberger, creator of Open Claw, discusses building an autonomous AI assistant that took the internet by storm, his journey from PSPDFKit to agentic engineering, and the future of AI agents.
要点 · TL;DR
与 AI 代理共情并引导它们是高效代理工程的关键。 Empathizing with AI agents and guiding them is key to effective agentic engineering.
像 Open Claw 这样的开源代理让非程序员也能编程。 Open source agents like Open Claw democratize programming for non-coders.
安全性，尤其是提示注入，仍是代理系统的主要挑战。 Security, especially prompt injection, remains a major challenge for agentic systems.
核心观点 · Key points
有效智能体工程的关键是共情智能体的视角并引导它。 The key to effective agentic engineering is empathizing with the agent's perspective and guiding it.
构建一个理解自身系统的 AI 智能体能够实现自我修改和强大的调试能力。 Building an AI agent that understands its own system enables self-modification and powerful debugging.
像 Open Claw 这样的开源 AI 智能体使编程民主化，允许非程序员做出贡献。 Open source AI agents like Open Claw democratize programming, allowing non-coders to contribute.
安全性是智能体系统的主要问题；提示注入仍然是一个未解决的问题。 Security is a major concern for agentic systems; prompt injection remains an open problem.
玩耍和实验对于掌握智能体工作流和建立直觉至关重要。 Playing and experimenting are essential to mastering agentic workflows and building intuition.
反共识 · Contrarian takes
永远不要回滚代码；而是让智能体向前修复，这样更快更高效。 Never revert code; instead, ask the agent to fix forward, which is faster and more efficient.
模型智能不会随时间退化；用户项目变得更杂乱，使智能体显得更差。 Model intelligence does not degrade over time; user projects become sloppier, making agents seem worse.
“氛围编程”是贬义词；严肃工作应称“智能体工程”，凌晨 3 点后才可氛围编程。 Vibe coding is a slur; prefer 'agentic engineering' for serious work, vibe coding only after 3 AM.
用语音提示智能体比打字更高效，即使对于复杂的软件开发也是如此。 Using voice to prompt agents is more efficient than typing, even for complex software development.
最适合智能体的编程语言不一定是你喜欢的；根据生态系统选择。 The best programming language for agents is not necessarily the one you like; choose based on ecosystem.
本期章节 · Chapters（共 63）
Open Claw 与 Peter Steinberger 简介Introduction to Open Claw and Peter Steinberger
PSPDFKit 起源与早期项目Origin of PSPDFKit and early projects
智能体的惊人自主性The Agent's Surprising Autonomy
开源与社区成长Open Source and Community Growth
智能体循环与迭代改进The Agentic Loop and Iterative Improvement
OpenClaw 为何胜出Why OpenClaw Won
智能体自我修改与影响Agentic Self-Modification and Impact
改名风波The Saga of the Name Change
规则与骚扰Rules and Harassment
改名压力Name Change Pressure
一切皆可出错Everything That Could Go Wrong Did
更名风波与商标问题Renaming saga and trademark issues
Molt Bot：艺术与病毒式反响Molt Bot: art and viral reaction
AI 误解与批判性思维AI Misconceptions and Critical Thinking
事件时机与公众反应Timing of the Incident and Public Reaction
安全挑战与 Open Claw BotSecurity Challenges and Open Claw Bot
提示注入与模型韧性Prompt Injection and Model Resilience
智能与攻击面的权衡Trade-off Between Intelligence and Attack Surface
安全问题与用户行为Security concerns and user behavior
与 AI 智能体合作需理解其局限Working with AI agents requires understanding their limitations
共情智能体视角Empathizing with the agent's perspective
放手，视智能体为团队成员Letting go and treating agents like team members
YOLO 方式与本地 CIYOLO approach and local CI
简短提示与语音输入Short prompts and voice input
对智能体的共情Empathy towards agents
实践的复合效应Compounding effect of practice
人在回路中 vs 全自动化Human in the loop vs full automation
文档与人性化Documentation and Human Touch
Soul.md 与 Anthropic 宪法Soul.md and Anthropic's Constitution
Soul.md 中个性的魔力The Magic of Personality in Soul.md
Soul.md 与记忆Soul.md and memory
开发者工作流与监控器Dev workflow and monitors
智能体作为开发者Agent as a developer
模型智能感知与退化Model intelligence perception and degradation
Claude Code 与 OpenClaw 对比Comparison of Claude Code and OpenClaw
操作系统与生态扩展Operating systems and ecosystem expansion
为智能体使用构建 TrimmyBuilding Trimmy for agentic use
OpenClaw 的 Mac 应用与打磨Mac app for OpenClaw and polish
苹果开发者体验与 AIApple's developer experience and AI
在不同硬件上运行 OpenClawRunning OpenClaw on different hardware
让 OpenClaw 人人可用Making OpenClaw accessible to everyone
给智能体 AI 初学者的建议Advice for beginners in agentic AI
用 AI 和开源学习编程Learning to code with AI and open source
AI 智能体时代最佳编程语言Best programming language for the AI agentic world
人生与成功指标建议Advice on life and success metrics
对 CEO 之路不感兴趣Not interested in CEO path
考虑大厂：Meta 与 OpenAIConsidering big labs: Meta and OpenAI
产品验证与 Anthropic 封禁用户Product validation and Anthropic blocking a user
赞美与结束语Compliment and closing thoughts
其他项目与未来计划Other projects and future plans
Meta 与 OpenAI 之间的抉择Decision between Meta and OpenAI
Open Claw 组件与主动智能体Open Claw components and proactive agent
热爱与项目原创性Love and Project Originality
模型心跳与共鸣Model Heartbeat and Relatability
MCP vs 技能与 CLI 方式MCP vs Skills and CLI Approach
Twitter CLI 与 API 访问Twitter CLI and API Access
AI 垃圾与人类价值AI slop and human value
智能体驱动的应用与服务交互Agent-driven interaction with apps and services
AI 取代程序员与编程未来AI replacing programmers and the future of coding