Fundamental world models for agentic vibe engineering

Agentic coding right now feels like gambling.One run, the agent lands it. Next run, garbage — sometimes in a programming language that doesn't exist.The disaster isn't the one-shot CLI. It's month three of a real project, when the agent has forgotten every decision you made and the only fix left is regenerating the whole codebase.The mistake: we keep treating code as the memory of the project.Code is a side effect. It doesn't capture intent. It doesn't explain design. It doesn't tell you why this seam exists and not that one. We hand the agent the side effect and call it memory — then act surprised when the design decays.What we actually need is a world model for the engineering project itself. Not a magic spec. A substrate.I've been writing about world models and memory for agentic AI for a while (books in comments). Here's the four-layer version specific to coding:→ Architecture and non-functional decisions — the shape of the system and the properties it must preserve. Latency budget