The AWS Machine Learning Blog published an article on October 8, 2026, about pay-per-inference for AI agents. It presents examples involving Incarna and BlockRun and describes payments for services at the level of individual requests. The central capability is Amazon Bedrock AgentCore payments, which the article describes as a managed way for agents to make payments. It handles payment protocols, wallet connections, transaction signing, and spending limits.

The practical idea is to include payment handling in an agent’s service request flow. The article describes a connection among the agent, a wallet, and a merchant providing a paid service. This flow involves more than displaying a price: it includes authorizing and signing a payment, then returning proof to the provider. In the examples, payments are tied to model inference requests, so each call can have its own price and settlement.

How Incarna configured wallets

Incarna provisioned a wallet for each agent using the Coinbase CDP connector. According to the article, the customer owns the wallet and grants Incarna delegated authorization to use it. This arrangement identifies who owns the wallet and who can perform payment actions under the granted authority. The article presents wallet setup as one part of integrating payments into agents.

AgentCore payments handles several elements needed for a payment: payment protocols, wallet connections, transaction signing, and spending limits. In the described system, a paid call needs to be associated with a wallet and operate within a spending boundary. The article presents these elements as a managed capability available for agent payment flows.

From an HTTP 402 response to payment proof

The article describes a service responding with HTTP 402. AgentCore payments can handle that response by making an x402 payment, signing it with the configured wallet, and returning cryptographic proof to the merchant. The merchant receives proof associated with the request being paid for. In this flow, the protocol, wallet, and signature connect a request for a paid service with confirmation of its payment.

Incarna used USDC on the Base network for settlement. AWS says each transaction is verifiable on-chain. That detail matters in the example because payments are associated with individual requests and need not be combined into one aggregate payment. The article describes both how a payment is initiated and how its settlement can be verified.

Two x402 schemes and spending boundaries

AgentCore payments supports two x402 schemes in the article’s account: exact for a known upfront price and upto for dynamic pricing. With exact, the price is known in advance. With upto, the agent authorizes a ceiling and the provider settles actual usage up to that ceiling. The distinction describes two pricing situations: one set before a request is performed, and another where the final amount depends on the service usage.

Incarna’s payment sessions had a spending ceiling and an expiry time. The company sized sessions to a day’s budget. In the article’s description, these settings bound the agent’s payment authorization: a session has a financial limit and expires after a set time. This shows how a limit can be applied over a working period rather than leaving a payment session without a defined boundary.

Integration timeline and beta activity

AWS reports that Incarna completed the AgentCore payments integration in three days: one day to build and two days to test. The application code was roughly 200 lines. The original estimate had been two to three months. These figures describe Incarna’s experience with this integration and give the article’s comparison between the initial estimate and the time the company reported after completing the work.

During beta, agents processed more than 1,000 payments. The amounts ranged from $0.001 to $0.05 per call, and each payment settled individually on-chain. The payment count and range illustrate the type of activity described: frequent, small payments associated with individual calls, rather than a single price for a batch of requests.

BlockRun and inference payments

The article describes BlockRun as a pay-as-you-go inference router serving model inference over x402. Each call is independently quoted, authorized, paid, and settled in USDC on Base. Alongside Incarna’s example, this describes a setting where an agent makes inference requests and the payment flow is repeated for individual calls.

AWS’s account connects several specific elements: the managed AgentCore payments capability, a customer-owned wallet used by Incarna under delegated authorization, handling of HTTP 402, two x402 schemes, and session limits. The Incarna and BlockRun examples cover request-by-request inference payments and settlement in USDC on Base. The article also gives an integration timeline, beta payment volume, and payment range. These details are presented as examples reported in the AWS article.