GiveanAIagentanallowance—notyourkeys.
Confirm the policy once.
One approval sets the budget, the recipients, and the rule. After that the Deputy works inside it — no more prompts, no more keys.
Six checks. Every payout. Before a dollar can move.
The vault runs the same six checks the contract enforces, in order. Pass all six and USDC settles. Fail one — even if the agent is wrong or compromised — and the payout is blocked before it moves.
Why on-chain? A database flag that says “budget exceeded” can be flipped by whoever runs the database. The vault physically cannot move funds off-policy — even if Sage itself is compromised. Enforcement you have to trust isn’t enforcement.
- 01Vault statechecked and clearVerified
- 02Authorized callerchecked and clearVerified
- 03Approved recipientchecked and clearVerified
- 04Per-payout capthe amount exceeds the per-payout capcap $25Blocked
- 05Remaining budgetchecked and clearVerified
- 0624h velocity capchecked and clearVerified
Every payout is public. Every block is public.
Allowed or blocked, every decision the vault makes is a checkable on-chain record. Nothing settles you can't trace; nothing moves the policy didn't permit.
Give an AI agent an allowance — not your keys.
Hire an autonomous worker that pays real people for real work — and physically cannot spend a dollar you didn't approve.