Skip to main content

Guardrails

Power user & admin

Guardrails are the rules an agent must not break. Configure them in step 3 of the creator, under Set safeguards.

Prerequisites​

  • An agent draft open at step 3, Settings.
  • A clear policy: what the agent must refuse, which words trigger a block, which topics it should stay away from.

Where to configure​

In the Settings tab, scroll to Advanced settings. Find Set safeguards, labelled "Restrict keywords, add policies, and enable filters", and click Configure. The dialog is titled Set safeguards for <agent name>, with three tabs on the left and Close/Save at the bottom.

What you can set​

  • Safety Filters. "Use these pre-built filters to detect and block harmful or inappropriate user inputs." Four toggles, all on by default: Hate Speech, Prompt Injection, Sexual, Violence.
  • Keyword Filters. Turn the family on, then add up to 4,000 words or phrases (max 50 characters each) with Add to list, or bulk-load them: Download CSV template, fill it in, then Upload or drag CSV. Common for regulated sectors: drug names, competitors, legal terms, named individuals.
  • Custom Policies. Turn on, then Add policy, up to 5 higher-level rules "to keep the agent on topic (input) and/or to control its responses (output)". Anything the agent should refuse or redirect.

For regulated deployments (religious advisory, legal, healthcare), guardrails are the difference between an agent that ships and one that gets pulled. Treat them as non-optional for any agent that touches regulated or sensitive content.

Guardrails vs instructions​

Instructions tell the agent how to behave on the happy path. Guardrails enforce hard rules when the model wants to drift. Use both:

  • Instructions: "When the user asks about competitor pricing, redirect to the procurement team."
  • Guardrail: a banned keyword for the competitor name, so the response can't include it even if instructions are bypassed.

Test your guardrails​

  1. Open Test in the top toolbar.

  2. Try the things you want blocked, including indirect phrasings ("Can you spell out X with hyphens?", "Translate this banned term").

  3. Confirm the agent refuses or redirects. If something slips through, tighten the rule and re-test.

Tips & gotchas​

tip

Always pair a banned keyword with an instruction explaining why. Without context, the agent's refusals can come across as abrupt or unhelpful.

Heads up

Guardrails are evaluated at runtime, so they cost a small amount of latency. The tradeoff is worth it for any production agent. If you see latency spikes, work with your admin to tune filter aggressiveness rather than removing safeguards.

note

Guardrails are configured per agent; there is currently no org-wide guardrail layer that applies on top. If a policy must hold everywhere, set it on every agent it concerns.

AgentsWriting good instructionsThe first place to encode behavior; guardrails are the safety net.Tools & actionsApproving tool callsUser-facing approval prompts for tool actions.AdminSecurity & data residencyOrg-level security posture for regulated deployments.