In plain language
They are the rules that stop an AI step from sharing private data, making things up, or taking actions it should not.
Example
Before sending an AI-drafted reply, the workflow blocks it if it mentions a refund amount above policy or includes another customer’s data.
Related terms
Prompt Injection · Human-in-the-Loop · Hallucination · Structured Output
Common misconception
A well-written prompt is not a full guardrail; important limits should also be enforced in code, permissions, and review steps.
Implementation guidance
Validate outputs against rules and schemas, restrict tool permissions, filter sensitive data, route uncertain results to people, and test with adversarial inputs.