When an Agent Should Stop and Ask: Three Clear Escalation Rules

举报
yd_287652226 发表于 2026/11/02 14:40:19 2026/11/02
【摘要】 Three rules for when an agent should pause and ask instead of guessing: before irreversible actions, when a required input is missing, and after repeated failure or conflicting evidence.

An agent that never asks a question will eventually guess wrong. An agent that asks about everything will waste the person’s time. The difference is not confidence. It is whether the next step can be undone, checked, or corrected cheaply.

Escalation is not failure. It is a planned pause: stop, state what is missing or uncertain, and ask for the one decision that unblocks the work. Three rules make that pause predictable, for the agent and for the person reviewing it.

1. Stop before an irreversible action

If an action cannot be undone - deleting data, sending a message to a customer, changing a production setting, spending money - do not infer the missing detail. Stop and ask, even if the request sounds urgent.

A useful ask names the action, the exact value that is uncertain, and the consequence of getting it wrong. For example: “I am about to delete the staging database named orders-test. Please confirm the name before I continue.” That is one decision, not a general question.

2. Stop when a required input is missing

Agents are good at filling gaps from context. That is also the risk. When a task needs a specific input - an account, a date, a file, an approval - and it is not in the request or in a trusted source, the agent should ask rather than invent it.

The test is simple: could a reasonable reviewer tell, from the request alone, which value to use? If not, ask. Record the answer with the run, so the next step does not have to ask again.

3. Stop after repeated failure or conflicting evidence

One failed attempt can be a transient error. Two failures with the same cause, or two sources that disagree, are a signal to pause. Trying the same call a third time rarely helps, and it can hide the real problem.

At this point the agent should report what it tried, what each source said, and what it needs: a corrected input, a different permission, or a human decision about which source to trust.

How to ask well

A good escalation is short. It states what was done so far, what is blocked, and the exact choice needed. It offers the options it can see, including “stop here” when stopping is a real option. It does not ask the person to redo the agent’s work.

Keep the record with the run: the question, the answer, and who gave it. That record is what makes the next review fast, and it is what lets the same rule work the next time without another conversation.

Stopping to ask is not a weakness in an agent. It is how an agent stays useful when the cost of being wrong is higher than the cost of waiting for one clear answer.

【声明】本内容来自华为云开发者社区博主,不代表华为云及华为云开发者社区的观点和立场。转载时必须标注文章的来源(华为云社区)、文章链接、文章作者等基本信息,否则作者和本社区有权追究责任。如果您发现本社区中有涉嫌抄袭的内容,欢迎发送邮件进行举报,并提供相关证据,一经查实,本社区将立刻删除涉嫌侵权内容,举报邮箱: cloudbbs@huaweicloud.com
  • 点赞
  • 收藏
  • 关注作者

评论(0)

0/1000
抱歉,系统识别当前为高风险访问,暂不支持该操作

全部回复

上滑加载中

设置昵称

在此一键设置昵称,即可参与社区互动!

*长度不超过10个汉字或20个英文字符,设置后3个月内不可修改。

*长度不超过10个汉字或20个英文字符,设置后3个月内不可修改。