Do people find these incredibly throwaway and short form posts that interesting to discuss? It comes off as lazy to me.

> Giving access to a precomputed entity map helps steer an agent

You can produce this from any relational database too though, especially if it's only for a few dozen entities.

I think the case for model discipline in LLM harnesses is huge, though. It's a worthy pursuit regardless of how you derive the model, from a graph database, mongo, sqlite, what have you.