AI agents are moving from answering questions to completing tasks, and that shift is now reaching ecommerce. A user asks an agent to find something, compare options, choose, pay, and return the result. In some cases that means buying API access. In others it means renewing software, ordering office supplies, booking a business trip, or making a low-value repeat purchase.
For merchants, this creates a practical problem. Online stores were built for humans — pages, buttons, forms, confirmation emails, support links. A software agent needs something different: structured product data, clear purchasing rules, a way to request or lock a price, an executable payment method, and a machine-readable confirmation that the order completed.
Forecasts for agent-led commerce vary widely, with estimates ranging from hundreds of billions to more than a trillion dollars by 2030. The exact number matters less than the direction. Agents are being prepared not only to search the web, but to transact on it.
Several large companies are building the rails. Coinbase is pushing x402, which lets websites and APIs request payment directly over HTTP. Google has introduced AP2, focused on authorization and accountability when agents pay. Stripe and OpenAI are working on agentic checkout inside ChatGPT. Visa and Mastercard are developing trusted-agent frameworks on existing card rails. MCP, the Model Context Protocol, gives agents a structured way to connect with external tools and services.
The pace has picked up sharply in recent weeks. On May 7, AWS launched Amazon Bedrock AgentCore Payments in partnership with Coinbase and Stripe, built on the x402 protocol and enabling AI agents to autonomously pay for APIs, MCP servers and web content using stablecoins. AWS also joined the x402 Foundation, with Warner Bros. Discovery among the early testers. Around the same time at Stripe Sessions 2026, Stripe extended its Agentic Commerce Suite to Google AI Mode and Gemini, introduced agent wallet infrastructure through its Privy subsidiary, and co-authored the Machine Payments Protocol with Tempo. AWS has said future phases will extend beyond digital resources into broader commerce, including flight bookings, hotel reservations and merchant payments.
These efforts are sometimes discussed as if one protocol will win. The more useful question is which protocol, payment rail or authorization model fits which use case.
A pay-per-use API call has little in common with a basket of physical goods. Machine-to-machine data purchases differ from family holiday bookings, and both differ again from a B2B agent renewing SaaS tools. The transaction values, refund risks, delivery methods and authorization requirements are different in each case.
x402 and stablecoins look well suited to API access, digital goods, machine-to-machine flows, microtransactions and cross-border services where instant settlement matters. Card-based agent payments fit mainstream retail, travel and consumer purchases, where existing fraud controls, dispute processes and customer protections already work. Mandate-based protocols come into their own when the key question is whether the agent was properly authorized to act on behalf of a user or business.
Most merchants will end up combining several of these rather than picking one. An agent might use MCP to discover a service, a mandate framework to prove authorization, x402 to make a stablecoin payment, and a merchant API to receive delivery confirmation. Another merchant may use card-network agent payments for checkout while still relying on MCP for product discovery. The mixed stack is probably how agentic commerce becomes workable in practice.
This week’s move from Cryptorefills, a digital commerce platform that has been processing stablecoin and digital asset payments for gift cards, mobile top-ups, eSIMs, flights and hotel stays since 2018, illustrates the point. The company announced x402 payments for AI agents at checkout — agents can now complete purchases using USDC on Base. That is the same x402 and USDC-on-Base stack AWS embedded into Bedrock AgentCore. Cryptorefills had already integrated MCP and Agent Skills, both of which already supported agent discovery and payment through its existing crypto rails. x402 adds a complementary, HTTP-native rail designed for autonomous machine-to-machine flows where the buyer is software and settlement is instant.
The quieter half of the same announcement, and arguably the more useful piece for other merchants, was an open-source agentic commerce repository covering how merchants can think about agent-facing commerce: conceptual guides, protocol comparisons, merchant playbooks, agent-facing flows and runnable TypeScript examples connected to live MCP and x402 endpoints. Documentation is released under CC0 and the example code under Apache 2.0.
The repository is useful beyond Cryptorefills because it was produced by a live merchant rather than a protocol team. While protocol teams focus on the technical specifications, this merchant-led open-source documentation bridges the gap to actual implementation, offering the kind of operational playbook that has been missing from the agentic commerce discussion. It is less concerned with proving that an agent can send a payment, and more with what a merchant has to build around that payment for the transaction to be usable — catalogue exposure, quote handling, price validity windows, late-payment behaviour, delivery confirmation, and the records that support, compliance and finance teams need afterwards.
Those are the questions merchants face before exposing checkout to agents. Which protocol fits the product. Which payment rail fits the authorization model. What the agent can see, what it can buy, how long a quote is valid, what happens when payment arrives late, how delivery is confirmed. A merchant-side reference connects the protocol discussion to those operational questions in a way the protocol releases themselves rarely do. The open-source documentation and example code from Cryptorefills is a publicly accessible reference for merchants considering how to structure their own agent-facing systems.
Agentic commerce will not advance only because large technology companies publish new standards. It will advance when merchants can decide which tools fit which use cases, combine them where useful, and expose their systems to agents without breaking checkout, fulfillment, support or reconciliation. The checkout page is becoming a set of instructions, permissions, payment options and fulfillment rules that software agents have to understand. The merchants that make those rules clear will shape the next phase.
