Write a policy once, then check a change against the graph before it ships — check_policies returns exactly what it'd violate. A query you run, not silent interception, and no re-explaining "don't touch prod" every prompt.
use it to
Check a change against your rules before it ships — see exactly what it'd violate.
Encode architectural rules — "mobile never calls the billing DB directly" — as a query you run against the graph.
Stop burning tokens re-explaining the same guardrails on every task.