Coding agents are excellent at local edits and weak at remembering the invariants that make a backend coherent.
When the only memory is chat context, every session re-discovers aggregates, event boundaries, and topology — you pay in tokens and review time.
Katip (Clerk MDA) keeps a versioned YAML model: Manifest → Domain → Architecture → Technology. Validation fails closed before generation.
This is not a claim that agents are useless. Pair an AI IDE for day-to-day coding with a model SSOT so architecture does not thrash.
Developer Preview honesty: we generate structural NestJS skeletons and strict validation — not a full hosted app platform.