Why AI Coding Agents Lose Engineering Context Between Sessions
AI coding agents can read and modify codebases effectively, but they often lack access to the reasoning behind engineering decisions made in previous sessions. A developer working on long-term projects observed that agents could see what a rule was but not why it existed, who approved it, or what depended on it. This creates risk when agents fill knowledge gaps with reasonable-sounding inferences that quietly become part of the codebase. The author argues that different types of engineering knowledge — product requirements, architectural decisions, verification results — need distinct, authoritative homes rather than being lumped into a single prompt or instruction file. The proposed solution involves structuring knowledge so that missing decisions surface as explicit gaps requiring human input before implementation begins, not after.
