智能体 / NAVIN RESEARCH
Human Approval Gates for High-Risk AI Agent Actions
Place approval at the last responsible moment and show people exactly what the agent intends to do.
Place approval at the last responsible moment and show people exactly what the agent intends to do.
简要回答: Require approval for irreversible, externally visible, expensive, or privilege-changing actions. Generate a preview from validated structured arguments, freeze those arguments while approval is pending, and execute exactly the approved operation.
实施步骤
- Classify tools by risk and define policy outside the prompt.
- Show target, scope, cost, data exposure, and rollback options in the approval view.
- Bind approval to a hash of the normalized request.
- Expire approvals and require re-approval after any material change.
验证清单
- Change an argument after preview and ensure execution is blocked.
- Test cancellation, timeout, and duplicate-submit behavior.
- Audit who approved what without recording excess sensitive content.
常见错误
- A generic 'Are you sure?' dialog.
- Approval before the agent has concrete arguments.
- Letting the model decide whether its own action requires review.
实际决策
先交付能够通过上述检查的最小设计。记录测试中使用的模型、提示词、工具、数据集与运行时版本，让后续变化可以被比较而不是被猜测。依赖供应商的限制、模型名称、价格或预览功能之前，请重新检查链接文档；这些细节变化往往快于本文的工程原则。
