Skip to main content
Guardrails are the deterministic rules that constrain what an agent will say and do — the difference between “usually behaves” and “provably won’t cross a line.” Unlike the agent’s prompt, guardrails are enforced, not suggested.

​
What guardrails cover

  • Topic and policy limits — things the agent must refuse or redirect.
  • Caller verification — require identity verification before disclosing sensitive information, so the agent never reads back an account detail to an unverified caller.
  • Escalation rules — when to hand off to a human.

​
In the API

Guardrails are part of the published version — publish to apply changes.