Hard blocks vs soft warnings: agent guardrails that work

Guardrails fail when they are vague. 'Be careful with customer data' is not a guardrail. 'Never send raw email addresses to a third-party API' is.

Make the lists explicit

• Hard blocks — the agent must refuse and escalate. No exceptions, no 'unless'.
• Soft warnings — the agent proceeds, but flags the action for review.
• Logged-only — the agent runs freely; behaviour is sampled later.

Write the hard list short. Every entry costs you flexibility. But every missing entry is a future incident.