An AI agent can put one API call on your AWS bill
AWS Marketplace now lets agents pay per request for select tools. The harder question is what they should be allowed to buy.
A research agent hits a paywall halfway through a job. It doesn't need a subscription. It needs one search. On November 19, AWS opened a way for agents to buy individual requests from select AWS Marketplace sellers and put the charge on an existing AWS bill.
AgentCore already supported agent payments before this release. The new part is how those Marketplace purchases settle: in dollars on the AWS account, without setting up a crypto wallet or opening a billing account with each seller. A participating endpoint sends back a price quote. AgentCore checks the payment session's budget, then issues a token the agent uses to retry the request.
That makes a paid lookup less of an interruption. It also means a request that used to stop at a paywall has become a spending decision. A cap answers how much the agent can spend. It doesn't tell the agent whether paying for a second lookup is useful, or whether a free source would have answered the question well enough.
I'd start with one approved seller, one task and a budget small enough to lose. Check the payment record against what the agent actually delivered. If the paid lookup didn't change the answer or save a person time, there's no reason to make it the default.
Sources
- 01 AWS Marketplace now supports agentic payments with AgentCoreAWS
November 19 announcement of per-request USD purchases from select x402-enabled Marketplace products on an existing AWS bill.
- 02 Enable AI agents to pay for AWS Marketplace x402 products with Amazon Bedrock AgentCore paymentsAWS Artificial Intelligence Blog
Distinguishes August payments availability from November AWS Billing integration; explains quotes, session budgets, tokens and settlement.
- 03 AI agent productsAWS Marketplace Buyer Guide
Background on Marketplace AI agents and tools, discovery, deployment and AgentCore integration.
Related Kryden News
When should an AI agent stop and ask first?
Google DeepMind, MIT Technology Review, and Gartner are all pointing at the same practical question: what should an AI helper be allowed to do, and when should it stop for a human yes?
Robot demos are only useful when failure is part of the loop
NVIDIA GEAR's ENPIRE research shows coding agents moving from code generation into robot-training loops, where memory, resets, verification, parallel evaluation, and failure repair matter as much as the base model.