Payments
AI agent payments, explained
How wallets, spending policies, payment challenges, stablecoins, fiat settlement, and receipts work when software agents become buyers.
TL;DR
AI agent payments are transactions initiated by software within authority delegated by a person or organization. A safe system separates the agent’s ability to request a purchase from the policy engine’s authority to approve and sign it.
What to remember
- The agent proposes a purchase; a policy-bound account authorizes it.
- A wallet is only one payment method. The control model matters more than the rail.
- Protocol payment evidence should be enriched with intent, provider, policy, and outcome context.
What happens in a paid request
- 01
The service states its terms
A payment challenge identifies the amount, currency or method, recipient, and the resource being purchased.
- 02
The account checks policy
The requested spend is evaluated before a signature, token, or other credential is issued.
- 03
The payment is authorized
The system creates proof through the selected protocol and payment method.
- 04
The service verifies and responds
The provider follows the selected method’s verification and settlement procedure, then returns the resource when payment requirements are satisfied.
- 05
The system records evidence
The protocol returns settlement evidence or an optional receipt; Weft can attach the task, provider, endpoint, policy decision, and measured result.
Stablecoins, cards, and processor balances
The current x402 reference ecosystem is strongly oriented toward stablecoin payments, while MPP’s core payment-authentication design is method and currency agnostic. Stripe’s specific MPP integration supports stablecoins settled onchain, plus fiat methods such as cards and buy-now-pay-later through Shared Payment Tokens.
Protocol, payment method, settlement network, and merchant payout are separate choices. Stripe says its MPP transactions can settle to a merchant’s existing Stripe balance in its default currency; that is a property of Stripe’s integration, not of MPP universally.
Controls teams need in production
- Pre-transaction policy checks with explicit allow and deny reasons.
- Idempotency so retries cannot accidentally create duplicate charges.
- Short-lived, least-privilege payment credentials.
- Provider, endpoint, currency, and protocol observability.
- Receipts that finance and engineering can reconcile independently.
- A kill switch and budget controls outside the agent’s prompt context.
Sources and further reading
Protocols change quickly. We reviewed the primary sources below on July 23, 2026.