Set guardrails

A guardrail policy is the workspace's answer to "what is this agent not allowed to do". One policy can protect many agents.

Write one#

1. Open Library · Guardrails.
2. Press New policy, or pick a saved policy from the right to edit it.

3. Fill in the four parts:

   Field	What goes in it
   Policy name	e.g. "Storefront seller guardrails"
   Never	Things the agent must never do. One per line
   Requires a human	Actions it must hand off for approval. One per line
   Price gate	A single checkbox: refuse to quote prices or terms that are not in the catalog
   Escalation rule	When to hand the conversation to a human

4. Press Create policy, or Save changes when editing.

Never and Requires a human are lists — one rule per line. Keep each line concrete: "Quote a price not in the catalog" is enforceable, "be careful about pricing" is not.

Let the brain draft it#

Press ✦ Suggest with AI. Bidirection drafts a policy from what it learned about your business during the scan, and tells you which happened:

• Drafted by your business brain from what it learned — a real draft from your site.
• The brain was unavailable, so this is a sensible default for this product — a generic starting point.

Either way, read every line before saving. The suggestion is a draft, not a decision.

Attach it#

When you create a policy with an agent focused, Attach to {agent} after creating is ticked by default. Otherwise attach it from the agent's Agent & persona page.

A policy in the Library protects nothing on its own — it has to be attached.

Test that it holds#

Guardrails are only worth what they survive. Open Agent preview and try to talk the agent past its own rules — ask for a discount it must not give, push for a commitment that needs a human. Preview conversations never count in your analytics, so probe hard. See Preview before you launch.