Killing Vibe Coding: The 5 Layers of the "Human-on-the-Loop" Developer
Vibe coding is the human-out-of-the-loop problem applied to software engineering. The failure is not the model. It is the absence of institutional knowledge in the coding loop.
Describe what you want, let the AI write it. At small scale this works. At organisational scale it produces amnesia, where the model forgets an instruction thirty minutes later; convention drift, where the same pattern is implemented four different ways across four files; and security blindness, where a key gets hardcoded because nothing in the session said it could not be.
The episode’s claim is that none of these are model defects. They are what happens when a capable model is given no institutional memory. COC answers with five layers: specialist agents carry intent, skills carry context, rules and hooks carry guardrails, project instructions and commands carry the workflow, and an observation loop carries learning forward between sessions.
Two consequences follow. Rules re-injected on every turn cannot be forgotten, which is the anti-amnesia mechanism rather than a promise about attention. And a single constraint gets enforced through several independent mechanisms, so the failure of any one of them is caught by the next.
Raw model capability is becoming a commodity. Institutional knowledge is not. The full treatment is in the CO specification.