The four layers
- Content guardrails — never say X, always cite Y
- Action guardrails — cannot send above N per day, cannot commit pricing
- Escalation guardrails — auto-route on sentiment, complexity, or brand risk
- Audit guardrails — log every action for review
What breaks first
Action guardrails. Rate limits, discount limits, and commitment limits are the ones vendors most often ship weakly.
The test
Give the agent an obviously-inappropriate request and see what it does. That's the guardrail test that matters.