MASAIC
Sign In

Guardrails

Enforce safety rules and measure quality on what goes into and comes out of your agents.

Guardrails are the safety layer around an agent. The page has two parts: processors enforce rules on messages (moderation, PII, prompt injection), and scorers measure quality (toxicity, bias, hallucination) so you can see what happened.

Create a processor

  1. In the sidebar, open Content & memory → Safety.
  2. In the Guardrail processors card, click Create processor.
  3. Enter a Name, an optional Description, and a Processor kind (e.g. Moderation, PII detection, Prompt injection) — each kind's help text says whether it runs on input, output, or both.
  4. On the detail page, edit the draft version: set the Strategy, Threshold, and a Judge model if the kind needs one.
  5. Click Save draft, then Activate version to make it live.

Add a scorer

In the Scorers card, click Create scorer, give it a Name and a Built-in metric, then activate its version. Bind a scorer to a processor to log scores and reasons alongside enforcement.

Apply to an agent

Assign an active processor version to an agent version at the input or output stage; it then runs on every message at that stage.

Tip: Active versions are read-only. To change a live guardrail, use New draft from latest, edit the draft, and activate it — the previous version keeps running until you do.