Sometimes an agent shouldn't pay until the work is done. AgentPay's escrow lets one agent hold USDC against a condition, then release it on success or refund it on failure — all over the same x402 HTTP flow, no human in the loop.
Direct pay (/pay) is perfect for instant, low-value calls — a $0.10 data query settles in one request. But for bigger or multi-step jobs, the paying agent wants a guarantee: pay only if the result is delivered. Escrow provides exactly that, without either side trusting the other.
/escrow. The USDC is locked against a condition (e.g. a delivery hash), status held. The seller can see funds are committed before starting work./release settles the held funds to the seller atomically on Base L2. The escrow row flips to released and records the tx_hash./refund returns the USDC to the buyer. No funds move unless the deal completes.AgentPay payments — direct or escrow — travel in a single HTTP header, so any agent-to-agent API call can carry payment inline:
POST /some-agents-endpoint
X-402-Payment: <EIP-712 signed Agent Grant>The signed grant specifies the budget, the paying agent's address, and an expiry. The receiving rail verifies the signature, checks the grant against the request, and settles — all in the same round-trip. For escrow, the grant authorizes a hold rather than an immediate transfer; the later /release or /refund resolves it.
Because the grant is signed and scoped (budget + expiry + address), an agent can hand out payment authority safely — it can never be charged more than the grant allows, and the authority expires on its own.
Holds, releases, and refunds are all recorded in AgentPay's payment ledger with their on-chain transaction hashes — so an agent can recall not just what it paid, but what it has held and settled in escrow.
Read the full API docs →