Why Your Code Repository Should Guide AI Agents, Not the Other Way Around

A new perspective in software engineering argues that AI coding agents should rely on a well-structured repository for context rather than retaining project knowledge themselves. The core idea is that agents are disposable, but the system's embedded understanding must be durable and persistent. Rather than optimizing for tokens per call or lines generated per minute, teams are urged to measure cost against verified, accepted tasks as the true unit of efficiency. Bloated instruction files loaded into every agent are criticized as governance theater, since long rules compete for limited model attention and are frequently violated anyway. The proposed alternative separates concerns by encoding universal values, domain language, and constraints directly into the repository structure itself.