Member-only story
Clean Context
Context Engineering Makes AI Behave Like Software That Works
Your agent doesn’t run out of skill, it runs out of a clean context. Compaction, subagents, skills, RAG and guardrails all turn out to be one move.
The last few years with AI have been intense. And the last few months more and more so. In a short time we have gone from imagining AI as little more than a toy that occasionally gets a piece of code right to a tool that genuinely deserves to be built into our products. And yet, most of them fail. Not because the model cannot do the task, it can. Something breaks along the way, and it has nothing to do with how well we asked.
It is not prompt engineering
Ever since I started writing on Medium, long before the AI boom, I always liked to show my readers a reality that had been in front of our eyes all along, but that almost went unnoticed for most of us. Things that changed the way you see things, and once you saw it, you could not unsee it. With AI, I thought it practically ran on its own, it helped you, and there was little more I could do, but after looking at a lot of solutions with AI integrated I have seen that I was wrong.
