AWS published an account on October 8 of an agent that buys model inference one request at a time, with the spending ceiling set outside the model. It is the second major vendor this week, after Google, to describe an agent spend cap enforced by the platform.
What AWS describes
According to AWS’s post, Incarna built agents that pay BlockRun for inference over the x402 payment protocol using Amazon Bedrock AgentCore payments. AWS says the integration took three days and about 200 lines of application code, against the two to three months Incarna had scoped. Across the beta, agents processed more than 1,000 payments from $0.001 to $0.05 per call.
The control is a payment session with a maximum spend and an expiry time. AWS says AgentCore payments checks each payment against that budget before signing it, and that the agent’s code and prompt cannot change the limit.
Where the cap sits differs by vendor
Google’s agent, covered in our earlier piece, pauses when a project reaches an administrator’s cap. AWS sets the budget per payment session, which Incarna sizes to a day’s budget. A project cap bounds a project’s total AI spend, while a session cap bounds what one runaway task can spend, and neither vendor’s description covers both.
What it means for the buying committee
Treat the two caps as separate line items in a vendor review. Ask who sets the session budget, who can raise it mid-task, and where payment logs land for audit. AWS says the service emits logs and traces to CloudWatch and AgentCore Observability. The figures above come from one customer’s beta, so request production numbers before committing, and compare the unit economics against the agent business case we examined on October 8.
Source: Amazon Web Services
