Human approval on an AI agent is not authorization: bind it to the payload
Summary
分析了人类对 AI 代理的审批为何可能实际授权了与人类所见不同的操作，并提出在执行时对关键字段重新计算摘要进行比对、以单次条件更新消费授权的解决方案。
Similar Articles
Not every agent action should need the same approval
作者探讨 AI Agent 审批流程的设计问题，主张审批应基于风险等级而非组织架构——读操作放行、常规更新记录放行、高风险操作才需指定人工审批，并指出真正的业务规则往往存在于未文档化的实践者经验中，而不是 API 权限模型里。
How are you actually deciding which agent actions need human approval before executing?
The article discusses the challenge of determining which AI agent actions require human approval, citing a $27M unauthorized transfer in January 2026, and proposes a framework based on reversibility and impact.
An approval should stop applying when the action it describes changes
提出一个针对文件发送型 AI agent 的合成验收测试：审批应绑定到具体收件人、内容修订与动作并带过期时间，字段任一变化都需重新授权，且应记录拒绝后的实际状态证据。
where do you draw the line on what you let an agent log into
作者探讨构建会操作真实账户的 AI agent 时的授权边界：读取类操作允许常驻访问，而发送、消费、删除或发布类操作每次都需人工确认，并询问社区如何在大规模运行 agent 时处理写权限凭证。
Human approval is not a weakness in AI agents
The article argues that human approval is a critical mechanism for building trust and defining policy in AI agents, rather than a weakness to be eliminated. It suggests using approval patterns to iteratively expand agent autonomy safely.