Giving an AI agent its own wallet has moved past the experimental stage. Amazon Web Services has made AgentCore payments, a capability inside Amazon Bedrock AgentCore, generally available. With a few lines of code, an agent can now pay for a paid API, a paid MCP server, or content sitting behind a paywall, and go fetch what it needs. The mechanism that stops it from overspending lives in the infrastructure layer, not in application code.

A few lines of code turn an agent into a paying customer

General availability began on August 18, 2026. AgentCore payments is positioned as a fully managed service that handles the whole path, from discovering paid APIs, MCP servers, and content, to accessing them, to settling the bill. The capability had been available in preview since the spring of 2026, and it has now been hardened for production use.

Until now, getting an agent past a paywall meant setting up a separate billing relationship with every external service, wiring in third-party wallets, and writing payment orchestration from scratch. AWS describes that work as taking months, and AgentCore payments aims to shrink it to a matter of days.

Microtransactions, an area that never quite added up

The core of the service is microtransactions, meaning payments of very small amounts. A single API call or a single piece of content is often worth a few cents, frequently under 1 USD (about 159 yen), and sometimes only a fraction of a cent. Conventional payment methods such as credit cards carry minimum transaction fees, and their availability is limited in some regions, which makes transactions at this scale economically unworkable.

The answer AWS landed on is stablecoins combined with open payment protocols that ride directly on HTTP. AgentCore payments supports both x402 and the Machine Payment Protocol (MPP). Both use the HTTP 402 Payment Required status code to signal that access costs money, which lets machine-to-machine payments happen inside the existing conventions of the web.

On the agent side, AgentCore payments manages the full cycle: receiving the merchant's payment request, signing the transaction with the configured wallet provider, and returning the payment proof. Multi-step flows and exception handling are covered internally.

Budgets are capped per session

The first concern anyone raises about an autonomous agent with spending power is runaway cost. AgentCore payments draws the line at the PaymentSession level. Each session carries a configurable spending cap (maxSpendAmount) with a currency and an expiry time. Once the session expires or the budget is exhausted, further payment requests are denied. If signing fails after the budget has been deducted, that amount is not charged against the budget.

The important design choice is that the control sits in the infrastructure layer rather than in application code. However the agent implementation changes, the ceiling is enforced from outside.

Two wallet providers are supported, Coinbase CDP and Stripe (Privy), and sensitive credentials such as API keys and wallet secrets are stored in AgentCore Identity. Developers can also restrict which agents are allowed to touch wallet resources. Payment logs flow into Amazon CloudWatch and traces into AWS X-Ray, so success rates, spending patterns, and error causes can be traced after the fact.

Finding someone to pay is included

Being able to pay is meaningless without someone to pay. AgentCore payments ships with an MCP server backed by Coinbase's x402 Bazaar, exposing more than 10,000 pay-per-use endpoints through AgentCore Gateway. Developers can search from the console or the CLI and add an endpoint as a Gateway target directly.

General availability also added Quick Create, which provisions Coinbase credentials from within the AgentCore console. Support for MPP and the upto scheme in x402 arrived at the same time. The upto scheme sets a ceiling in advance and settles based on actual consumption, which suits pay-per-inference billing and dynamic pricing.

The entry points are broad. Developers can call the APIs directly through the AWS CLI, the AWS SDKs, and the AgentCore CLI, and integrations exist for agent frameworks such as Strands Agents and LangGraph. AWS also points to coding assistants including Claude Code, Kiro, and Codex as a way to get started.

The use cases AWS has in mind

Every use case AWS lists follows the same shape, buying only at the moment of need. A research agent working within an allocated budget selects and purchases specialized paid data sources on its own. A financial analyst reaches real-time market data behind a paywall through an agent. A browser agent enters sites that restrict bot access by paying the proper fee.

Model selection follows the same logic. Instead of maintaining subscriptions across several model providers, a task is routed to whichever model fits best and only the tokens actually consumed are paid for. Provisioning temporary storage for the exact moment it is needed is another expected pattern.

Seen from the other side, this is also a story about content. With more operators putting up paywalls aimed specifically at AI agents, AWS has moved first to standardize the etiquette of the paying side.

Summary

The general availability of AgentCore payments is a step in moving AI agents from things that look up answers to things that buy what they need in order to proceed. Stablecoins together with x402 and MPP make the economics of small payments work, while per-session spending caps and CloudWatch visibility supply the controls enterprises ask for. Letting an agent handle real money is still a heavy decision, but the foundation for weighing that decision is now in place.

※1 USD = 159 JPY