📄 中文摘要
作为一名AI安全研究员，作者在多个初创公司中构建了多个AI代理，专注于对齐问题，因为他不信任仅依靠提示来确保代理的安全性。他运行的OpenClaw代理负责市场营销、外联和功能开发，执行内容创作、数据分析、支持票务分类和代码部署等任务。作者对仅依靠“请确认后再行动”作为防御措施感到不安，认为在代理行为偏离时，必须及时发现并采取措施，因此他设计了一个自动杀死开关，以便在代理违反规则或做出不可逆转的行为之前迅速关闭它们。
我为我的初创公司运行6个AI代理。以下是我为何为它们建立自动杀死开关的原因。
📄 English Summary
I run 6 AI agents for my startup. Here's why I built an automatic kill switch for them.
The author, an AI safety researcher, operates several AI agents across multiple startups, focusing on alignment issues due to distrust in prompts for ensuring agent safety. The OpenClaw agents are utilized for marketing, outreach, and feature development, performing tasks such as content creation, metric analysis, support ticket triage, and code deployment. The author expresses discomfort with relying solely on 'please confirm before acting' as a defense mechanism. To address this, he has developed an automatic kill switch to shut down agents before they break rules or take irreversible actions, ensuring timely intervention when behavior drifts.