Governance for agentic AI requires access control at every level – from tool discovery through query execution to response synthesis – not just at a single central checkpoint like in RAG.