Security & trust
Evidence for auditors, not assertions.
The audit stream is structured for compliance queries — actor, policy version, approver, outcome — not narrative claims.
The permit layer
- Authorization boundary Agents call authorize() before acting. No valid permit, no execution.
- Signed permit ES256 JWT scoped to workspace, action type, and resource reference.
- Policy engine Define allow, approval threshold, and deny rules per action type.
- Approval flows High-risk requests route to human approvers. Permit issued on approval.
By action type
- Refunds & Credits The primary wedge. Permit-gated before they execute.
- Customer Commitments Promises your agent can't make without authorization.
- Account & Plan Changes Destructive changes that need more than a flag check.
- Pricing Overrides Discounts your agents can't self-authorize.
- Production Actions Config changes, deploys, and database ops with a permit trail.
Start where it would hurt.
Pick the use case where an unauthorized action would cost you — refund, cancellation, or pricing change — and add the permit boundary in under an hour.
Start with refundsVeritrellis vs
Guides
Controls & trust