Weekly Digest 7
AI
Context Engineering
Context Engineering is a quite popular term in the past weeks, and I have read a lot articles talking about it. Here I made a summary introducing these articles:
一文看懂“提示词” vs “提示词工程” vs “上下文工程”,
Author: Baoyu
提示词工程是一个过程，系统化地设计、测试、优化提示词的过程。
上下文工程（Context Engineering），就是一门为 AI 设计和构建动态上下文的学科，为大语言模型提供恰当的信息和工具，帮助模型高效完成任务。
对于普通人来说，能写提示词就够了，要开发 AI 应用才需要考虑提示词工程去不断优化提示词，要开发动态的 AI 智能体才需要去搞上下文工程为 AI 的上下文窗口填充恰好的信息。
The New Skill in AI is Not Prompting, It’s Context Engineering
what is included in context engineering:
- Instructions / System Prompt
- User Prompt
- State / History (short-term Memory)
- Long-Term Memory
- Retrieved Information (RAG - Retrieval Augmented Generation)
- Available Tools
- Structured Output
Karpathy’s tweet about context engineering
art because of the guiding intuition around LLM psychology of people spirits.
So context engineering is just one small piece of an emerging thick layer of non-trivial software that coordinates individual LLM calls (and a lot more) into full LLM apps. The term “ChatGPT wrapper” is tired and really, really wrong.
Context Engineering for Agents
How to manage Context
- Write Context
- Select Context
- Compressing Context
- Isolating Context
Agentic AI
SaaS vs. Libraries: from humanware to machineware
AI agents are becoming the primary uses of out digital infrastructure. Based on this idea, here is what the article discussed:
- Libraries are becoming preferred architecture over api
- Libraries grant agents complete control.
- MCPs that are not build natively for AI are not good for use.
- What the emerging agent-native stack looks like in practice.
- How this shift fundamentally changes the role of engineers are the nature of software itself.
- Machines become the customers of products.