Mandate lets you give an AI agent a budget instead of your keys. You fund a mandate with a spending cap, an expiry and an optional merchant allowlist. Every payment the agent proposes waits in a veto window on Arbitrum, and until it closes you can reverse it with one transaction.
AI agents are starting to shop, book travel, buy compute and pay for software. The payment options they have today are all bad:
A hot wallet is all or nothing. One prompt injection or leaked key drains everything in it.
Approving every purchase kills autonomy. At that point the human is the agent and the agent is a typist.
On-chain payments are final. Cards have chargebacks. Stablecoins do not. Nobody has built the missing grace period for software that spends on your behalf.
Mandate is a non-custodial payment vault on Arbitrum. An owner funds a mandate for one agent with hard limits: a total budget, a per-purchase cap, an expiry, a veto window and, optionally, an allowlist of merchants. The agent never holds the money. It can only call spend, which the contract validates, reserves funds for and places in a veto window.
Inside the window the owner can veto the payment (funds return to the mandate), approve it early, rotate a leaked agent key, or revoke the mandate and get the free balance refunded. If the owner does nothing, anyone can settle the payment once the window closes, so merchants never depend on the owner being online. The interface also flags payments that deserve a second look: first-time merchants, amounts near the cap, bursts of activity and merchants vetoed before.
The worst case for a compromised agent is bounded by the budget, the cap and a single window.
Every wallet, agent framework and operator putting money in an agent's hands has the same question: how much can I trust this thing? Mandate answers it with a number, a window and an undo button. It is a new financial primitive for agentic commerce: reversible, bounded, auditable spending authority.
The design only works if a payment, a veto and a settlement each cost a fraction of a cent, and if budgets are held in stablecoins an agent can actually spend. Arbitrum's low fees, EVM tooling and stablecoin ecosystem (including USDG) make that practical.
Limits live on-chain, not in a system prompt. A prompt is a suggestion; a contract is a rule.
It is the only model here where a payment can be undone before it settles, without giving anyone custody.
Settlement is permissionless and every payment carries a memo with the agent's stated reason, giving a built-in audit trail.
No admin keys and no upgrade path on the vault.