AI agent authorization
Authorization answers a specific question: may this identified agent perform this action, for this purpose, under these conditions? It is narrower than general access and should be traceable to a person, policy or other accountable authority.
Authorization is more than a prompt
A natural-language request can express intent, but it does not automatically establish identity, permission or legal authority. A robust design evaluates the request against an explicit scope and relevant policy before an external action is allowed.
The authorization record should make clear which subject requested the action, which resource is affected, what limits apply, how long the permission lasts and what evidence supports the decision.
Design for least authority
Grant only the access needed for a defined task. Narrow scope by action, resource, counterparty, amount or time where those dimensions are relevant; require stronger review when the impact or uncertainty is higher.
- Prefer a bounded, expiring mandate over an open-ended grant.
- Separate policy evaluation from payment or execution.
- Require escalation when context is missing or contradictory.
- Test denial, timeout, retry and revocation paths as carefully as approval.
What to verify
Ask where authorization is enforced and whether the enforcing system can be independently tested. The mandate guide covers the grant itself; spending controls covers one possible limit.