Solution
Approvals for agent actions.
Decide each call against your policy, ask a person where it matters, and be able to stop an agent quickly.
01
Policy decides every call.
- Write rules over action, risk class, assurance level, mandate, time windows and more. Each rule permits or forbids; the first match wins.
- Policies are versioned. Publish one, roll back to any earlier version, or start from one of four templates.
- A refusal names the policy hash and the rule that matched, so an agent, your support team and your auditor trace the same decision.
- Run the real evaluator on your machine with agined dev, or replay recorded calls against a draft to see which would change.
02
A person signs off.
- Mark an action as needing approval. The call waits with HTTP 412 while Agined sends the request to an approver over your notification channels.
- The decision is signed, and then the call goes on or stops.
- Name who may approve, let an approver delegate authority for a set time, and set an answer-time target per action. A missed target is flagged on your chain.
03
Authority that only narrows.
- A mandate gives an agent standing authority with limits on actions, spend and expiry, checked on each call.
- A sub-mandate can only be narrower than its parent. Revoke one mandate and every child stops at once.
- Sensitive platform actions need a single-use confirmation token that is burned when the call runs.
04
Stopping an agent.
- Kill switches for an actor, an action, a domain or a whole surface. Before you flip one, Agined counts what it would have refused over the last hour.
- Quarantine an actor and each of its calls waits for a person's approval.
- A per-action breaker trips when calls or errors spike, and each further call needs approval until it lapses after an hour or you clear it.
- Every flip is recorded on the chain, with who did it.