Guardrails and approval
Screen messages for personal data and jailbreaks, and pause a run until a person approves.
Guardrails
A Guardrails step inspects the message and follows Pass or Fail. It checks for:
- Personal data, so sensitive details do not reach a model or a tool by accident
- Jailbreak attempts, where a message tries to talk the model out of its instructions
Put it directly after Start, and connect Fail to a step that explains why the request was not handled, or straight to End.
Human approval
A Human approval step pauses the run until a person approves or rejects it. The run then follows the matching path.
Use it before actions that cannot be undone, such as sending a message, issuing a refund, or changing a record.
Approval on a tool call
A Tool call step can require approval on its own. The run pauses just before the call, shows what is about to be sent, and continues only if it is approved. This avoids adding a separate approval step.
A run waiting for approval stays paused. No later step runs until someone answers.