The reason these projects collapse isn’t “vibe-coding.” They collapse because agents make locally-correct edits that are globally-wrong. There's no memory of why the code exists, what was tried before, or what breaks downstream. Better rails need: > Structured memory: dependency graphs, call chains, ownership boundaries that a person can understand > Retrieval aware of blast radius, not just semantic similarity > Execution-backed verification (tests, builds) not text confidence Once agents can ingest an entire codebase state, simulate changes safely, and prove correctness with running artifacts, “knowing how to code” stops being the hold-up. "Knowing what to build" - under what constraints, with what tradeoffs - becomes the question. The moat shifts from typing to judgment and review. We'll get there - it's a matter of when, not if.