Today, the agentic economy has stopped being a prediction and become operational infrastructure. Giants like Shopify are already using protocols such as ACP and UCP to let AI agents buy in their stores. As a result, Shopify has recorded more than a million merchants selling directly inside ChatGPT, all thanks to the OpenAI and Stripe alliance, through the ACP (Agentic Commerce Protocol).
But it is not the only player here, Universal Commerce Protocol from Google and Shopify, launched with the backing of Walmart, Target, Carrefour, Zalando and more than twenty other global retailers, has also made its mark.
After all, the convergence of agent-to-agent communication protocols (Anthropic's MCP, Google's A2A, and IBM's Agent Communication Protocol) under the Linux Foundation umbrella has allowed AI agents to be no longer mere assistants: they discover products, negotiate prices, execute payments and fulfill contracts. They are autonomous economic actors. And the rails they operate on are being laid right now.
It was born to dethrone Python and is already open source: this is Mojo 1.0
The problem of the two ACPs and why the confusion matters
But before continuing, let's pause for a moment. At this point you have surely noticed, or may know, that the acronym "ACP" today designates two different protocols that solve different problems. The name collision is not coincidental and should be understood:
ACP (IBM/BeeAI) = Agent Communication Protocol
Launched in March 2025 by IBM Research as part of the BeeAI project, this protocol solves agent-to-agentcommunication: it allows an agent built with LangChain to talk to one from CrewAI and another developed bespoke, without glue code for each pair.
It includes cryptographic capability tokens that limit what each agent can do, and native hooks for OpenTelemetry (traces, metrics, logs). In April 2025 it was donated to the Linux Foundation AI & Data together with BeeAI Platform and Docling; in August 2025, IBM announced its merger with Google's Agent2Agent Protocol (A2A) under the same foundation.
The original GitHub repository is archived; migration to A2A is underway with A2AServer and A2AAgent adapters already available in BeeAI Framework.
ACP (OpenAI/Stripe) = Agentic Commerce Protocol
Presented on September 29, 2025, it standardizes agentic checkout: the connection between a buying agent (ChatGPT, initially) and the merchant. Its specification uses date-based versioning and defines composite blocks: agentic checkout (create, update, retrieve, complete, cancel sessions), cart and feed (browse catalogs, manage pre-checkout carts), delegated payments (secure tokens between buyer, agent and business), delegated authentication (OAuth 2.0 for agents to act on the user's behalf), orders and webhooks.
Version 2026-01-30 introduced the Payment Handlers Framework (breaking change: it replaces simple identifiers with structured handlers with ID, version, spec URL, PCI requirements, PSP identification), Capability Negotiation (intersection of features between agent and seller) and the Extensions Framework (discount codes as the first extension). It powers "Instant Checkout" in ChatGPT: the user discovers and decides in the chat, but the transaction is executed against the merchant's infrastructure.
The distinction is clear. The first ACP connects agents to each other; the second connects agents to commerce. A realistic scenario in August 2026: a personal agent (under A2A, heir to IBM's ACP) negotiates with a supplier agent (also A2A), which in turn invokes the merchant's checkout via ACP (OpenAI/Stripe) or UCP. The full stack requires both.
Microsoft's 18 rules for human-AI interaction
UCP, the "full journey" standard from Google and Shopify
But another important player is missing here, one that makes its entrance on January 11, 2026. That day, Google and Shopify present the Universal Commerce Protocol (UCP) with a coalition that aims to redraw the e-commerce map: Shopify, Etsy, Wayfair, Target, Walmart, Best Buy, The Home Depot, Macy's, Flipkart, Zalando, Carrefour, Adyen, American Express, Mastercard, Visa and Stripe join forces to change everything.
The architectural difference with ACP (OpenAI) is scope. UCP standardizes the entire agentic shopping journey: product discovery → cart → checkout → payment → after-sales (returns, support, loyalty). ACP (OpenAI), by contrast, focuses only on checkout.
Thus, UCP is not just a more complete tool, but one that follows the agent throughout the process so that there are no errors. Picture it as that sales assistant who kindly walks with you, showing you and talking about the merchandise you are interested in buying, except that UCP is a sales assistant for AI agents.
And that is why UCP is built on REST transports, A2A (Google's agent-to-agent communication) and MCP (access to tools and data), and integrates AP2 (Agent Payments Protocol) for the payment layer. The protocol is of course open source and you can find it here. There are no walls, everything is free, a big difference from Stripe and OpenAI's ACP, which is tied and locked to their platform.
In fact, Stripe has been smart. It is backing both standards: it is co-creator of ACP (OpenAI) and guarantor of UCP. May the best one win.
DFlash 2: the technology that lets AI respond three times faster
Fitting all the pieces together
Of course, to navigate this ecosystem, it helps to see it as layers stacked on top of each other (bottom-up):
- Layer 1: MCP (Model Context Protocol), from Anthropic, November 2024. It connects models with tools and data (model-to-tools). By December 2025 it had more than 10,000 servers in production. Adopted by Claude, ChatGPT, Microsoft Copilot, Google Gemini, VS Code. Donated to the Linux Foundation in December 2025, giving rise to the Agentic AI Foundation. It uses JSON-RPC over stdio/HTTP. It is the tool that allows an agent to invoke the merchant's inventory API, CRM or ERP without integration for each tool.
- Layer 2: A2A (Agent2Agent Protocol), from Google, April 2025. It connects agents with agents (agent-to-agent). It now absorbs IBM's ACP under the Linux Foundation. It defines discovery via Agent Cards (declarative metadata: name, description, capabilities, endpoints, skills), runtime capability negotiation, and execution of long-running tasks with streaming updates. It is the nervous system of the multi-agent economy. BeeAI Platform, previously driven by ACP (IBM), already uses A2A natively.
- Layer 3: WebMCP, W3C standard, February 10, 2026. It brings the MCP model to the browser via
navigator.modelContext. Proposed by Google and Microsoft engineers within the Web Machine Learning Community Group. In June 2026 it entered public origin trial in Chrome 149, aimed at agents operating inside the browser (Gemini in Chrome). A store declares invocable tools (catalog search, cart management, order lookup) and the agent calls them directly. The key architectural difference: with WebMCP, the website itself acts as an MCP server; tools are defined and executed in the browser, without additional backend infrastructure, complementing traditional MCP servers.
- Layer 4: UCP / ACP (Google/OpenAI/Stripe). UCP covers the full journey; ACP (OpenAI) covers checkout. Both use MCP/A2A/AP2 underneath. Stripe acts as the universal connective tissue: its Shared Payment Token (SPT) and Delegated Payments Spec work with both protocols. If a merchant already processes with Stripe, it can enable agentic payments by updating a single line of code; if it uses another PSP, it can keep using it and add Stripe only for the agentic layer.
The mental diagram is very clear: MCP (backend) → WebMCP (browser) → A2A (agent↔agent) → UCP/ACP (commerce). Each layer solves a different part of the journey; a merchant can adopt them in parallel. The groundwork is the same, a structured catalog (prices, stock, shipping, policies) that feeds Schema.org, Merchant Center feeds and, now, WebMCP tools.
An economic shift without precedent
And all of this together leads to an economic shift unlike anything seen before. Because if today every person interacts with online commerce platforms directly, in the agentic scenario every person will be able to activate hundreds or thousands of agents to execute tasks simultaneously. Basically, a professional sleeps while their agents negotiate contracts, analyze data, buy supplies or solve operational problems. On waking, the work is already done.
Of course, we are still very far from that. Today's AIs cannot work 100% autonomously and reliably. But with the speed of advances in the sector, that goal is not far off in the future. A reality that could completely change global commerce.
Think about it for a moment. We are not talking about millions of users; we are talking about billions of active agents consuming services, demanding infrastructure, making transactions and generating data in unprecedented volumes. Consumption becomes exponential.
Swarm behavior
After all, agents do not behave like humans: they operate 24/7, make decisions in milliseconds, consume APIs at scale and demand real-time responses. Like a swarm of bees working in unison at the command and disposal of the queen. Here that would be the AI, for the humans who control it.
That is why those who adapt early to this new reality on the way will build the foundations of this new economy; those who delay will have to run to catch up. The competitive opportunity window is open now.
Companies that understand this movement will have not only a technological but a strategic advantage. And here the question stops being "How do we serve the user better?", and becomes "How do we operate in a world where the user may not be human?". Success will depend on your answering this too.
Why AIs will end up funding the journalism they consume