The main concern I have is that there's not a well defined security context in any agentic system. They are assumed to be "good" but that's not good enough.