I've spent the past several months building an agentic AI platform, and the thing that ate the most time wasn't prompting or orchestration. It was permissions. A realistic setup has an agent framework, a BI tool, and something for notebooks. Each has its own notion of who may read what. So a rule as simple as "this agent can only see these three tables" gets implemented three times, in three shapes, and drifts apart the moment someone changes one. Nobody notices until an agen...