Guardrails & allowed models
Guardrails are the limits AI operates within; allowed models control which AI models it can use. Together they let an admin set a safe envelope that everything in the organization runs inside.
Organization-wide guardrails
Section titled “Organization-wide guardrails”Individual agents have their own guardrails, but an admin can set org-wide guardrails that apply on top of every one of them — a floor no agent or user can drop below. So even if someone builds an agent and forgets a limit, the organization’s guardrails still hold. This is the safety net that makes broad, self-serve building safe.
Allowed models
Section titled “Allowed models”Rival is model-agnostic — work can run on different AI models. Admins decide which models are permitted, and can steer work toward the right one for the job. That control serves several goals at once: capability (use a strong model where it matters), cost (use a cheaper model where it doesn’t — see Cost controls), and data policy (restrict where sensitive work can run).
Configuring guardrails
Section titled “Configuring guardrails”Admins set guardrails centrally, and they apply across the organization. Because they’re enforced by the platform rather than by convention, they hold consistently — no depending on each builder to get every limit right.
A quick example
Section titled “A quick example”An admin allows two models: a strong one for general work and a cheaper one for high-volume tasks, and sets an org-wide guardrail that no agent may send data to an unapproved connector. A team then builds freely — every agent they create is automatically inside those bounds.