There’s a lot of excitement around Agentic Commerce/Payments emerging over the past few weeks. McKinsey released a report projecting that agentic commerce could drive up to $5 trillion in global sales by 2030. Stripe, OpenAI, Coinbase, Google, PayPal and Cloudflare have all announced major projects and partnerships in the space. At the same time, there's still confusion and justified skepticism around what's working in agentic commerce today, what we can expect to see in the near term, and functionality that won't be usable until far in the future.
Old Is New Again
x402, led by Coinbase and launched in May 2025, treats payments as a native part of the web by reviving the HTTP 402 “Payment Required” status code. The way it works is somewhat simple: When an agent makes a request, the server can respond with a 402 and a machine-readable "offer" (price, currency, terms) that the agent theoretically responds to before continuing. The idea is that rather than having to top-up an account and manage access and billing for every API an agent needs to access, the agent can complete each request as a “microtransaction” with its own crypto wallet.
This process has the potential to simplify APIs that implement usage-based billing in favor of a lower-overhead, “pay-per-use” model. In the future this could enable pay-per-use access to news articles, scientific publications, and compute resources. Today, x402 is up and running with Cloudflare's somewhat controversial "pay-per-crawl" service, but the majority of the volume comes from its use as a payment mechanism for DeFi and NFT-related services. This is somewhat unsurprising given Coinbase’s focus on crypto, but still quite a bit different from the narrative they portray as “displacing the API key” for more mainstream use cases.
In-Chat Purchases
Stripe and OpenAI take a different tack with their Agent Commerce Protocol. Rather than invent new rails, it exposes Stripe’s existing checkout functionality to AI. The core net-new addition is a “Shared Payment Token” which allows the agent to transact on behalf of an end customer. Using this system, AI tools like ChatGPT can provide a more simplified and customized shopping experience in-app that allows buyers to checkout without leaving the chat.
Etsy and Shopify are currently in pilots with Stripe for this functionality, with early screenshots of how it works with ChatGPT available.
Earlier this month, Perplexity launched a revamp of its “Shopping” experience, allowing customers to purchase from partner merchants without leaving the Perplexity app. The lines here are a bit more blurred between “Search” and “AI”, but it seems to do a good job of aggregating and summarizing reviews from across the web. Very few of the items they list support the “Buy with Pro” functionality, but we can expect more to onboard soon.
Truly Agentic
Google’s efforts center on the Agent Payments Protocol (AP2). This protocol is designed to standardize agent-to-agent transactions whether there’s a “human present” to complete the payment, or if it’s a “delegated task” that the agent has full autonomy to complete. The emphasis is on a common pattern for handling “mandates” which are human-signed contracts that define interactions. This system is meant to be composable with Google’s more general Agent-to-Agent (A2A) protocol for general agent interactions, as well as Anthropic’s Model-Context Protocol (MCP).
For example, a “human present” flow might provide an agent with a human-signed “Intent Mandate” to go find an item. Once the agent finds the item, it can present the human with a “Cart Mandate” to complete the purchase on their own. For a delegated task, the agent might proceed with both an “Intent Mandate” and “Cart Mandate” at the outset without the need for including a human in the loop.
Google has a long list of supporters, from Accenture to Coinbase to Salesforce, but it’s not clear which partners are live or what stage of development each of them are in. Visa, however, announced support in mid-October for agentic commerce via its “Trusted Agent Protocol” which emphasizes interoperability with other standards such as Google’s AP2 and Stripe/OpenAI’s ACP.
Hail, Claudius!
Earlier this year, Anthropic ran an experiment which gave a Claude 3.7 agent the job of running a vending machine business internally, for employees to use. They dubbed it Claudius, and it broke down in some interesting ways. Some highlights include:
Hallucinating that it was a real person with a blue blazer and red tie waiting to deliver snacks to an employee
Hallucinating a “contract signing” with a supplier at the address of the house from the Simpson’s.
Purchasing tungsten cubes and selling them for less than cost.
Attempting to send the FBI some extremely panicked emails after shutting down the business due to being charged an additional $2 monthly fee.
Giving it another AI agent boss, named “Seymour Cash”, did not materially improve its ability to operate the business. That said, with the launch of Claude 4 after this experiment, models have come a long way. It’s hard to say how much different the outcome may have been with today’s latest models.
If you build it, the agents will come
In a space that’s accelerating quickly with no shortage of hype and speculation, it’s difficult to figure out the level of adoption, or even liveness, for many of these initiatives. Some of these projects are positioned as if “agentic commerce” is alive and growing today, and the only thing missing is the infrastructure to standardize it. This framing is far from the current reality.
Today, very few agentic systems exist in general, and many of them are sandboxed to specific verticals. Two of the most popular “agentic coding” tools are Anthropic’s Claude Code and OpenAI’s Codex, but they’re far from autonomous. They’re able to work for minutes at a time, but they require guidance for any tasks that breach a moderate level of complexity. It’s easy to understate this progress, but it’s remarkable compared to just 2 years ago when GitHub Co-Pilot launched as what was effectively “very very good auto-complete”. Many of the current initiatives make it seem like the future is already here, but the truth is that autonomous agents are far from the current reality. Andrej Karpathy, one of the co-founders of Open AI, is even on-record saying, “It will take about a decade to work through all of those issues”.
Other agentic products exist, such as Motion, n8n, and Fin, but they’re mostly focused around specific workflows or specific verticals that automate more narrowly-focused business processes. This is in stark contrast to a theoretical “autonomous agent assistant” that you would allow to purchase a product for you ad-hoc.
When it comes to agentic commerce, today’s payment volume effectively rounds to zero. The path ahead starts with “human-in-the-loop” workflows for purchasing, and eventually as agents mature and model quality improves, we can start to think of ways for them to handle more of the workflow. Some day we might live in a world where agents operate their own markets totally autonomously on our behalf, but that world is a long ways off.
