Military field manuals offer an unexpected blueprint for safer AI agents, from pre-execution red teaming to structured mid-run corrections.