护栏Guardrails

限制模型能做什么、不能做什么，防止它跑偏或做危险操作。Constrain what the model may and may not do, so it does not drift or take dangerous actions.

在输入和输出两侧加的一层检查规则，可以是关键词、权限、也可以是另一个模型判断。

A layer of checks on the input and output side — keywords, permissions, or another model's judgement.

给 Agent 配了写文件和跑命令的能力，就必须配护栏。这是能力与风险的直接交换。If you give an agent the ability to write files and run commands, you must give it guardrails. This is a direct trade of capability for risk.

OpenHands 的受限模式OpenHands Restricted Mode ↗

回到名词地图看第3站的其他 36 个词 →Back to the map: the other 36 terms in station 3 →

• OpenHands 的受限模式OpenHands Restricted Mode ↗