Trigger
An agent proposes a consequential action through an SDK, tool wrapper, gateway, or API adapter.06
Action authorization
Guard the one-way door in any agent workflow.
Let AI prepare the work, then put a deterministic or human decision before the refund, invoice approval, payout, booking, account change, or other irreversible commit.
What breaks
The final answer looks reasonable, but an intermediate tool call crosses a policy boundary, lacks trusted evidence, or exceeds the authority the principal granted.Control point
Create a typed proposed action, evaluate authority and evidence, then authorize, hold for approval, or deny before the existing system executes it.01Proposed action
→02Policy + evidence
→03Decision
→04Existing system
How it works
- 01Describe
Make the action explicit
Send the operation, amount, principal, counterparty, grant, evidence references, and idempotency key as a typed action.
- 02Decide
Evaluate facts outside the prompt
Deterministic policy checks the action against caps, spend windows, required evidence, known counterparties, and approval thresholds.
- 03Escalate
Hold the ambiguous cases
A held action does not execute. It waits for the named approver, and the approval or denial becomes part of the record.
- 04Prove
Join decision to execution
The decision receipt explains what was allowed. The execution receipt records what actually moved after authorization.
Other control points