I came to the same conclusion and built something to solve this pain point on top of having scoped access and tasks, it’s called [Takibi Base](https://takibibase.com/), a shared knowledge base for AI agents: one human operator keeps source documents, Profiles control what each agent can read, and agents retrieve exact cited passages; Takibi does not generate answers because there’s no LLM at all, you can just use whichever LLM you want instead.