Is it safe to let an agent spend money on your behalf?

The question behind the question

"Can I trust this agent with money" is usually the wrong frame. An autonomous agent is software following instructions, and the useful question is narrower: what is the worst thing a single signature can do? That is a property of the payment scheme, not of the agent's judgment.

Two models, very different blast radius

Token approval. The common pattern in on-chain applications. You approve a contract to move up to some amount of a token, often an unlimited amount, and it draws against that allowance whenever it likes. The approval persists until revoked. If the counterparty is compromised at any point after you approve, the allowance is still there. The blast radius is everything you approved, for as long as it stands.

Signed per-payment authorization. The agent signs one authorization per payment, covering an exact amount to an exact recipient, valid only inside a time window, and carrying a nonce that can be used once. Nothing persists. There is no standing allowance to revoke, because none was ever granted. The blast radius is one payment.

Both look like "the agent can pay." Only one bounds the loss.

What to check before letting an agent pay any API

• Does it request an allowance? If the flow starts with an approval transaction, the answer is a standing grant. Ask why it needs one.
• Is the amount fixed at signing time? The signature should commit to the exact number, not a ceiling the recipient chooses within.
• Is the recipient fixed at signing time? An authorization that lets the payee be substituted is not bounded.
• Does it expire? A signature with no validity window is a bearer instrument with no end date.
• Can it be replayed? A single-use nonce, enforced on-chain, is what stops the same authorization settling twice.
• Who pays gas? If the agent has to hold native currency to submit transactions, you are funding a second balance and a second failure mode.
• What happens on a duplicate request? Retries are normal in agent systems. Ask what the service does when the same work is requested twice.

How this service answers those

A retrieval is paid with an EIP-3009 authorization signed as typed data: exact amount, exact recipient, a validBefore window, and a single-use nonce. No token approval is requested or required, so there is no allowance to revoke afterwards. The signature is submitted by the payment facilitator, not by the paying agent, so the agent never needs a gas balance.

Before anything is reserved or settled, the paying address is screened against the OFAC Specially Designated Nationals list; a match is refused unpaid. A pointer settles at most once — a second paid request against one already settling is refused rather than charged again. The details are on the security page and in the sanctions compliance statement.

The honest limitation

If settlement succeeds but the item has already expired or been retrieved, the payment stands and the response is a not-found. That is a real edge, it is documented rather than hidden, and the Refund and Dispute Policy covers it.

Try it without an account

curl -X POST https://slidingbox.ai/v1/key

Returns an evaluation key immediately — no signup, no email, no approval step. Send it as X-API-Key on a retrieval to skip payment while the key has allowance left. Storing is free either way.