Four Agent self-check items that do not require rewriting the entire system and can be implemented this week: list the tools with write/send/submit/delete capabilities and mark the worst consequences; check the provenance labeling and authenticity verification along the path of content from generation to submission; set mandatory human confirmation rather than an optional item for actions toward external organizations or irreversible actions; and re-examine whether the task prompt describes the target state or the permitted behavior boundaries.