What did AWS report?
In its article on pay-per-inference, AWS describes a production payment flow involving Incarna, BlockRun and Amazon Bedrock AgentCore payments. Incarna’s agents pay BlockRun for individual model requests. Amazon Bedrock AgentCore payments handles payment authorization and enforces spending limits.
According to AWS, BlockRun serves more than 90 models from more than 15 providers through the x402 payment protocol. Each request is quoted and settled separately, without requiring a subscription for each provider.
AWS reports that Incarna completed the integration in three days, compared with an original scope of two to three months. Across the beta, agents processed over 1,000 payments ranging from $0.001 to $0.05 per call. These are results from the reported implementation, not a forecast for another business.
How does an agent pay without approving its own budget?
AWS explains that a paid endpoint returns an HTTP 402 payment request. The agent then uses AgentCore payments to authorize payment from its configured wallet. The seller verifies the payment signature before providing the inference. In the Incarna implementation, payments settle in USDC on the Base network.
The key control is outside the model. According to AWS, a payment session has a spending ceiling and an expiry time. AgentCore payments enforces that ceiling, and the agent’s code or prompt cannot change it. Incarna sets its sessions around a day’s budget.
AWS also describes two payment schemes. The exact scheme is typically used when the price is known beforehand. The upto scheme lets an agent authorize a maximum amount, with the provider settling for actual usage up to that limit.
What does this mean for a small or medium business?
Paying by request can make spending follow actual activity more closely. That may suit a workflow with uneven demand. But a small charge is not the same as a small total bill. Repeated calls, retries or an agent stuck in a loop can still consume a budget.
The useful business distinction is between permission to perform a task and permission to spend money. A model should not be the only safeguard for either. Set spending authority through controls that the agent cannot rewrite, and assign a person to own the budget.
A payment ceiling also does not guarantee a useful result. Review task quality alongside cost. For businesses in Malaysia and elsewhere, finance teams should examine funding, reconciliation and accounting requirements before approving a wallet-based payment workflow.
What should you do next?
Start with one task that genuinely needs paid services during execution. Set a maximum spend per task and a time limit. Define when the agent must stop or ask for human help. AWS describes testing with testnet USDC; use a test environment before committing real funds.
Before a production pilot, confirm who owns the wallet, who grants signing permission and who can revoke access. Review payment records against completed work. Track total spend, failed requests and useful outputs rather than judging the pilot only by its per-call price. Use the AWS source article to review the described setup.
Separately, if your priority is delegating work rather than building payment infrastructure, Terabot gives businesses digital workers: AI workers with their own private virtual machines that work in the apps a team already uses and operate around the clock. It is in private preview, by invitation only.
Key takeaways
- AWS reports a working flow in which Incarna agents pay BlockRun for individual model requests through Amazon Bedrock AgentCore payments.
- Infrastructure-enforced spending limits matter more than instructions telling an agent not to overspend.
- Pilot one workflow and measure both useful output and total cost before increasing spending authority.
Written with AI assistance from the source linked above, and checked against it before publishing. Product names belong to their owners. Check the original source before relying on details.