Technical and policy boundaries that keep AI agents safe, legal, and aligned with organisational values.

AI guardrails include: abort conditions on every Play, compliance checks embedded in execution workflows (like the WCAG example), data sovereignty rules (what data can agents access, where can it flow), cost limits (maximum spend per agent per loop), output quality gates (automated evaluation before anything reaches users), and ethical boundaries (what agents are not permitted to do regardless of instructions). The Integrity Steward owns the guardrail framework, and Guard agents enforce it in real time.

Guardrails should be versioned and reviewed as the system's autonomy expands.