TLDRocket
Sign in

Pay-per-inference for AI agents: How BlockRun and Incarna use Amazon Bedrock AgentCore payments

Amazon Web Services Peter Jiang

AWS showed AI agents paying for model calls one by one. It cut Incarna’s setup from months to days and kept spending capped.

Based on reporting by Amazon Web Services, Peter Jiang — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Amazon Bedrock AgentCore payments is AWS’s new answer to a very specific problem: how do you let an AI agent buy tiny things without handing it a credit card and hoping for the best? In this case, the “tiny things” are model calls. Incarna used the system so its agents could pay BlockRun for inference one request at a time.

That sounds niche until you look at the economics. These are sub-cent purchases, repeated over and over inside an agent’s loop. Card rails were never built for that. So AWS pushed the job into infrastructure: AgentCore payments handles the protocol, connects to a wallet, signs the transaction, and enforces spending limits outside the model itself.

The BlockRun setup is built around x402, the payment protocol used for each request. BlockRun acts as a pay-as-you-go router for more than 90 models from more than 15 providers, with each call quoted and settled separately. When a paid endpoint returns HTTP 402, the agent can respond through AgentCore payments, sign with the configured wallet, and hand back cryptographic proof to the seller.

Incarna used the Coinbase CDP connector to provision each agent wallet, while the customer kept ownership and granted delegated authorization. Payments settle in USDC on Base, and every transaction is verifiable on-chain. The spending controls live at the platform layer, which means a manipulated prompt does not get to rewrite the budget mid-flight. AWS says Incarna sized sessions to a day’s budget.

The headline result is the part builders will care about: Incarna says the integration took three days, about 200 lines of application code, instead of the two to three months it had originally scoped. Across the beta, agents processed more than 1,000 payments ranging from $0.001 to $0.05 per call, with each settlement happening individually on-chain.

My take — AI-written commentary, not fact-checked reporting

This is the right shape of automation: let the agent spend, but only inside a box the platform controls. The industry has spent years pretending “autonomous” and “unlimited” are the same thing; they’re not, and the wallet should not be a vibes-based decision. Also, anything that turns “three months” into “three days” is going to get attention whether the market is ready or not.

Read more about this at: Amazon Web Services

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.