
An unmarked server module sits on a pull-out tray in a bright aisle of rack-mounted cloud hardware. AI-generated editorial illustration.
An AI agent can now buy an eligible tool from AWS Marketplace and put the cost on its developer’s AWS bill. Amazon added the option to Bedrock AgentCore payments. Teams can restrict how much the agent can spend and which sellers it can pay.
Amazon calls this an extension of the payments feature it made generally available in August. The November addition is the Marketplace link. It lets developers put some small tool purchases on the same bill as the rest of their AWS use.
What an agent can buy in AWS Marketplace
The new route covers eligible software, APIs, data, and content that support x402, a protocol that lets a service ask for payment when software tries to use it. In its November 19 launch explanation, AWS says those payments are in US dollars and appear on the AWS bill. They can work without a stablecoin wallet or separate API key for the purchase.
Take a team that asks an agent to check addresses for a set of stores. A data provider might charge each time the agent makes a request. The team could let the agent use that provider for a small job, within a limit. That is the kind of purchase this route is meant to support.
The Marketplace route also gives agents a way to find eligible products. A team can connect the catalog to the tools its agent can reach. The services it can discover and the sellers it can pay are distinct decisions.
What the spending controls cover
AWS says teams can set a spending limit for a payment session and choose approved sellers. Once the session hits its cap, the agent cannot keep spending through that session. The company says teams can trace transactions through CloudTrail and review the costs in Cost Explorer.
Those controls address different risks. A limit can hold down the size of an unexpected bill. An approved seller list narrows who gets paid. A transaction record lets someone inspect what was bought after a run. Teams still have to pick settings that fit the task they are handing over.
For example, if a failed request leads the agent to try again, it helps to know what the seller charges and what the agent logs. A small cost per call can add up in a job with many steps. The team can check those costs alongside how often the tool was useful.
Useful details to check include when a new session starts, who is able to raise a cap, and whether a paid request sends data to a seller. Our guide to AI agent permissions covers the general choices behind that review.
Why billing belongs in the agent design
AWS’s AgentCore payment documentation also covers other payment routes. The news here is the ability to bill eligible Marketplace use in dollars. A team can now put an expense on its cloud bill while an agent carries out part of a job.
That makes the bill another place where the system’s actions show up. Before a team gives an agent a wider budget, it can compare the paid calls with the work they helped finish. Our guide to AI agent architecture shows where tools, state, and human review fit into an agent that can take action.
Read also
Popular news

Human clues helped AI agents mimic supply-chain decisions

Cline-Bench tests coding agents three times on real software tasks

ElevenLabs moves to buy Kotoba for its Japanese voice AI

Claude Code vs Cursor: Try the Same Fix in Both

US House lawmakers launch a bipartisan caucus on AI safety



