Guardrails

What a guardrail is

A guardrail is a safety valve: anything that stands between your AI and your user and can stop bad content from being delivered. On Future AGI, guardrails are a group of evals put to that job, user-facing safety checks like toxicity, sexism, and personal-information detection, run inline so the verdict becomes an action (deliver or block) instead of a stored score.

%%{init: {"flowchart": {"curve": "linear"}}}%%
flowchart LR
  A["Your agent's<br/>response"] --> G["Guardrail<br/>a safety eval, run inline"]
  G --> P["Passes<br/>delivered to your user"]
  G --> B["Fails<br/>blocked before delivery"]
  style G fill:#2f2f2f,stroke:#ffffff,stroke-width:2px

What makes an eval a guardrail isn’t the template, it’s the job: deciding in real time whether content ships.

Two ways to run guardrails

Through the gateway

Agent Command Center’s gateway guards traffic at the platform level. It ships built-in checks (rule-based and AI-powered), integrates third-party guardrailing systems, and offers Future AGI’s Protect models natively. You group checks into policies, choose enforce or monitor mode, and scope them globally, per project, or per key. Nothing in your application code changes; the gateway sits on the request path.

Through the Protect SDK

Protect is the in-code path: a small set of guardrailing evals built for millisecond latency, available only through the SDK’s guardrails module. Nothing is on by default; you enable the checks you want at the point where your code produces output, so the screen runs inside your application without adding noticeable turnaround.

Guardrails and standard evals

The same eval can hold both jobs. Toxicity is available as a guardrail (inline, blocking) and as a standard eval on the platform, where evaluator models like the Turing family give you deeper analysis over a dataset or your traces. Scoring yesterday’s responses for toxicity is evaluation; refusing to deliver today’s toxic response is a guardrail. If you don’t need live blocking, run the same checks as standard evals and skip the gateway entirely.

Keep exploring