How do you design guardrails for safe and predictable agent behavior?