Guardrails
Guardrails are the rules and checks that constrain what an AI is allowed to do, keeping it inside safe or intended boundaries even when a prompt or a model's own judgment might push past them.
An AI that can take real actions needs limits that don't depend on the model behaving well every time. Guardrails are how a system stays safe by construction — permissions, approval gates, and locks — rather than by hoping the AI always makes the right call.
Assistant
Every agent acts inside the same permission tree as a person, and anything that would gate on a human decision — an irreversible action, a governance change, or an agent asking for help mid-run — goes through the same Approval primitive: propose, then whoever already had the permission decides.