I kept hitting the same wall with LangGraph: tutorials show you how to build a graph, not how to maintain one when you have 8 nodes, 3 agents, and shared state across subgraphs.

So I built a reference architecture with: - Platform layer separation (framework-independent core) - Contract validation on every state mutation - 110 tests including architecture boundary enforcement - Patterns that AI coding agents can't accidentally break

Repo: https://github.com/cleverhoods/sagecompass Wrote about the patterns: https://dev.to/cleverhoods/from-prompt-to-platform-architect...

It's MIT licensed. Would love feedback on the approach - especially from anyone who's scaled LangGraph past the tutorial stage.