You ship a working feature on Monday. Tuesday morning a small cosmetic bug appears. You ask the AI to fix it. The fix works but the login flow now throws a 500. You paste the stack trace back in. The login comes back, but the payment webhook is silent. Eight prompts later, the UI is half-broken in three new places and you can no longer remember which version of the app actually worked. This is not bad luck, and it is not a skill issue. It is the predictable output of four compounding technical limitations in how today's AI coding agents understand code. By the time a founder is 30 prompts deep and watching previously-working features vanish, the tool has quietly lost the thread: the context window is full, the generation is non-deterministic, the agent has no map of what depends on what, and it has been patching symptoms rather than causes. The loop is escapable. The way out is architectural, not another prompt.…