Context Engineering is All You NEED!

I believe what is more important than context engineering is understanding how to properly break a problem down into sub systems or sub tasks.

then your name is context engineering

I watched the LangChain video regarding this and comparing to your assessment. I like that you've added the common issues in context engineering. When I'm using Perplexity, for example, I manually manage context to produce better results as the chat history grows alongside iterative prompting or rewriting/regenerating responses. But also, I'm deleting specific parts of the chat history which also helps in generating quality responses and preventing reaching context limit fast. I apply "manual context engineering" in both coding and non-coding use cases for tasks that require high accuracy results. When I take this practice of "manual context engineering" into agentic coding to produce the best results and/or fixing problems, this helps to alleviate the amount of debugging I have to do and reducing context poisoning which can metastasize if not kept in check. When I'm creating in-house MCP tools for my app and for my agentic coding tool, I'm applying the lessons learned from "manual context engineering" and applying into my tools and my app. Its a bit daunting and sometimes have to put aside these ideas to get work done!😆

Context engineering makes sense to me because prompt assumes we are doing a single prompt whereas context is a wider term that can involve multiple round trips plus additional contextual information from external sources, available tools, etc.

RAG Beyond Basics Course: https://prompt-s-site.thinkific.com/courses/rag

I think more in terms of context cultivation and context harvesting. Some system messages explain context cultivation and include a line that says "if I ask you to harvest, it means we have built sufficient context, review chat history to date and save a copy of of your understanding to desktop". I think the harvesting component is essential. A large portion of the context window is used engineering context, taking the engagement closer to degeneration. Harvesting and restarting with the AI's debloated interpretation means there's interpretation visibility, which can be edited if slightly misaligned and densified for greater per token potency. Doing this sets the utility benefit point (UBP) early, where latent space is more open, less contaminated, and tuned more precisely to the specific need.

pretty cool approach, thanks for sharing

This was really good session

Been diving into context engineering lately. AgentVoice’s way of keeping track of customer interactions makes so much sense now.

Intelligent and gracefully explained