The Pulse
AWS Shows AI Agents Paying for Inference Through AgentCore
AWS detailed how Incarna uses Amazon Bedrock AgentCore payments to pay BlockRun for AI model inference one request at a time. The company says its beta agents have processed more than 1,000 on-chain payments, with spending limits enforced o
AI.info Team · Written with AI ·
Incarna’s AI agents have processed more than 1,000 payments for model inference, each settling individually on-chain, AWS said Thursday. The transactions ranged from $0.001 to $0.05 apiece, illustrating how Amazon Bedrock AgentCore payments can handle purchases too small and frequent for conventional subscription billing.
The October 8 case study describes an existing service in use, not a new launch: AWS announced general availability for AgentCore payments on August 18. The latest account focuses on Incarna, an AI-agent product built by SpreadX, and BlockRun, an inference router that charges per model call.
From an HTTP 402 to a paid model call
When an Incarna agent requests inference, BlockRun returns an HTTP 402 “Payment Required” response with a quote for that call. AgentCore payments checks the amount against the session’s spending ceiling, signs the payment through the configured wallet and returns proof that BlockRun can verify before serving the inference.
BlockRun routes requests across a catalog of more than 90 models from over 15 providers. Each call is quoted and settled separately over the x402 payment protocol, so the agent pays for the inference it uses rather than maintaining separate provider subscriptions. AWS says the system also supports an “upto” option: an agent authorizes a maximum for a request, and the provider settles the actual usage up to that limit.
Spending limits sit outside the model
Incarna provisions wallets through Coinbase’s CDP connector. The customer owns the wallet and delegates authorization to Incarna; the payment itself settles in USDC on the Base network. AWS says each transaction can be verified on-chain, while AgentCore enforces session caps and expiry times at the infrastructure layer rather than relying on the agent’s code or prompt to follow a budget.
That separation matters because an agent can misread instructions or repeat a request. A session cap gives the system a hard ceiling that the model cannot change, according to AWS. Incarna sets session lengths to fit a day’s budget, the post says.
Three days instead of months
AWS says the Incarna team completed the integration in three days—one day to build and two to test—using roughly 200 lines of application code. The team had originally scoped the work at two to three months. Those figures come from AWS’s customer account; the post does not provide an independent comparison of implementation time or payment-system costs.
“AgentCore payments covered everything we needed for an agent to pay over x402: a wallet the customer owns, a funding and revocation flow, a spending limit the platform enforces, and signing that handles both versions of x402. We wrote none of it,” Justin Zhou, founder of Incarna, said.
A customer example for a service launched in August
AWS describes AgentCore payments as a managed capability that connects agents to wallets, handles payment protocols and enforces spending rules. Its August general-availability announcement said the service supports Coinbase and Stripe Privy wallets, and added support for the Machine Payment Protocol alongside x402.
The Incarna and BlockRun example shows one working arrangement: the agent initiates a request, a service sets a price, and infrastructure applies a spending limit before money moves. The reported beta volume establishes that the integration has processed real transactions; AWS’s post does not disclose the total value paid or provide usage figures beyond the 1,000-plus payments and per-call range.