Love this post! Was just reading this related post about “Context engineering by hand” and they seem to be related in terms of the stack they propose and the “skyscraper” here. It’s interesting that they mention and include MCP as a layer at the lower level. : https://open.substack.com/pub/aibyhand/p/context-engineering-by-hand
Adding a link to a related paper from Google about Context Engineering. A few things that, by my read, they miss:
In their paper, Google treats context as one monolithic "payload" rather than structured layers. All context types are bundled together without independent management: "Context Engineering addresses the entire payload, dynamically constructing a state-aware prompt based on the user, conversation history, and external data."
On page 7, they talk about "Strategically selecting, summarizing, and injecting different types of information" This seems to assume that optimization is applied uniformly across the entire context. Your skyscraper idea describes a better way.
Here’s another good Substack resource on context engineering. I liked one comment that predicts companies will be soon be dedicated to Context, and here we have the Substack for it!
Love this post! Was just reading this related post about “Context engineering by hand” and they seem to be related in terms of the stack they propose and the “skyscraper” here. It’s interesting that they mention and include MCP as a layer at the lower level. : https://open.substack.com/pub/aibyhand/p/context-engineering-by-hand
Great post!
Wonderful article Mark - truly
enlightening
Adding a link to a related paper from Google about Context Engineering. A few things that, by my read, they miss:
In their paper, Google treats context as one monolithic "payload" rather than structured layers. All context types are bundled together without independent management: "Context Engineering addresses the entire payload, dynamically constructing a state-aware prompt based on the user, conversation history, and external data."
On page 7, they talk about "Strategically selecting, summarizing, and injecting different types of information" This seems to assume that optimization is applied uniformly across the entire context. Your skyscraper idea describes a better way.
Here's the full paper on Kaggle:
https://www.kaggle.com/whitepaper-context-engineering-sessions-and-memory
Here’s another good Substack resource on context engineering. I liked one comment that predicts companies will be soon be dedicated to Context, and here we have the Substack for it!
https://open.substack.com/pub/addyo/p/context-engineering-bringing-engineering?r=3su05&utm_medium=ios