2025 年大家都忙着搞 agent. 下面分类是随便分的.
Tools/Prompts
- System Prompts and Models of AI Tools. 各种 AI 的系统提示词以及 tool schema
- Claude Cookbooks. citation 和 research agent 的提示词示范
- 2025-05. Highlights from the Claude 4 system prompt. 分析提示词
- 2025-03. Markdown vs. XML in LLM Prompts: A Comparative Analysis
- Claude Docs. Essential tips for long context prompts
- 2025-04. How ChatGPT Memory Works. 逆向 memory tool
- 2025-06. 逆向 Gemini 2.5 Pro 搜索功能. Browse 工具是个 sub-agent, 根据提示词从网页中返回相关信息给主 agent.
- Anthropic. 2025-03. The “think” tool: Enabling Claude to stop and think in complex tool use situations
- Anthropic. 2025-09. Writing effective tools for agents — with agents
Context-Engineering
- Lance’s Blog. 2025-06. Context Engineering for Agents
- Drew Breunig. 2025-06. How Long Contexts Fail
- manus. 2025-07. Context Engineering for AI Agents: Lessons from Building Manus
- 周星星-知乎. 2025-09. Context Engineering 上下文工程的前世今生
- 2025-10. Context Engineering for AI Agents with LangChain and Manus - YouTube. PMF 之前别训模型
- Anthropic. 2025-10. Introducing Claude Skills. 不同于之前的 tools/mcp, skills 可以层级化地提供信息, 提供信息的方式更灵活, LLM 按需加载, 而无需一开始就把所有 desc 都加载到 prompt 中. 上面那篇 manus 里也有提到他们做了类似的按需加载的模式.
System
现在趋势是自己不做 index (分块 + 向量化 + 向量数据库), 直接让 LLM grep 或者 web search. 比如 Claude, Cline, manus 都是如此.
- Cline. 2025-05. Why Cline Doesn’t Index Your Codebase (And Why That’s a Good Thing). Cline 有很多关于模型各种数据的博客.
- minusx. 2025-08. What makes Claude Code so damn good (and how to recreate that magic in your agent)!?. 大道至简
- Multi-Agent
- Anthropic. 2025-06. How we built our multi-agent research system
- Cognition. 2025-06. Don’t Build Multi-Agents
Deep Research
- Google. 2025-06. Gemini Fullstack LangGraph Quickstart. 最基本的 agentic search pattern
- LangChain. 2025-07. Open Deep Research. 结构和很清晰基础, 在搜索阶段分 topic 给 sub-agent 干活, 最后用一个 LLM 写报告. 比 Gemini Quickstart 多了最开始的澄清和计划步骤, 类似 Gemini 和 ChatGPT 中实际的 deep research. 也是基于 LangGraph 写的, 代码见 这里.
- Jina. 2025-02. DeepSearch/DeepResearch 实施实用指南. Jina 擅长做 embedding, 技术博客里有很多和 RAG 相关的文章, 风格相比别家也比较 tech.
- 周星星-知乎. 2025-04. 端到端的训练, 怎么复现 Deep ReSearch
Training: Agentic RL