Approve an agent's action (CIBA + RAR)
An AI agent asks for approval to do one exact thing — you approve it, and it gets a token for nothing more.
- 1
On the CIBA lab, pick one of your registered agents and describe a sensitive action (e.g. transfer $250).
The agent authenticates with its own bearer secret (through the SAME authenticateAgent kill switch as ai5) and asks to act on ITS OWNER — never anyone else — describing exactly what it wants to do as an RFC 9396 Rich Authorization Request.
- 2
Review the pending request and Approve it — you may be asked to step up first.
Approving a sensitive out-of-band request is gated exactly like rotating a client secret: a TOTP-enrolled approver must re-prove a fresh second factor before it goes through.
- 3
Poll as the agent and see the token it receives.
The minted token’s authorization_details claim carries EXACTLY what you approved, byte for byte — not a general-purpose credential, a token scoped to that one action. The pending request can only ever be redeemed once.
Learn the theory
🩻 X-ray — what actually happened
Your own insert-only audit trail — the real server events, sanitized (never a secret), each linked to the lesson that explains it.