For most of the web's history, software could read a price but it could not pay one. A program could scrape a checkout page, parse the total, even fill the form. The last step, handing over money on its own, was always a human's job. That assumption is now breaking. AI agents are beginning to transact, and a real payments stack is forming underneath them.
This is a field guide to that stack. It is deliberately neutral: who is building what, which protocols matter, and how the pieces fit. If your brand sells anything, these are the rails your next customer may arrive on, even if that customer is not a person.
Why agents need their own money rails
The payment methods we built for people do not fit machines. Cards assume a human approving a purchase. Logins assume a session and a password manager. Monthly subscriptions assume you can predict, in advance, which services you will use. An autonomous agent breaks all three. It may call a hundred different APIs in an afternoon, each for a fraction of a cent, none of them known when it started the task.
Two problems fall out of this. The first is settlement: agents need to pay small amounts, often, without a checkout flow built for hands and eyes. The second is identity. Before a merchant accepts money from a piece of software, it wants to know which agent this is, who it acts for, and what it is allowed to spend. Most of the activity in the space is an answer to one of those two questions, or both.
The protocol layer
Before the companies, the standards. A handful of protocols are competing to become the default way an agent and a server agree on payment.
| Protocol | What it does | Backed by |
|---|---|---|
| x402 | Revives the dormant HTTP 402 "Payment Required" status so a server can demand a stablecoin payment in-line, settled on a chain like Base. No account, no signup. | Coinbase (open source) |
| ACP | Agentic Commerce Protocol. A standard for an agent to run a full commerce flow against a merchant, from product to checkout to payment. | Stripe, OpenAI |
| MPP | Machine Payments Protocol. An internet-native way for one machine to pay another, sitting alongside x402 as a payments primitive. | Industry consortium |
| MCP / UCP | Not payment rails themselves. MCP handles tool discovery, and UCP handles checkout, the layers an agent crosses just before it pays. | Anthropic and others |
These are not mutually exclusive. A single transaction might use MCP to find a tool, ACP or UCP to assemble the order, and x402 or MPP to move the money. The stack is still settling, and the names will change. The shape, discovery then checkout then settlement, is already clear.
The players
Pay-per-call access and wallets
AgentCash is a clean example of the wallet end of the stack. Its pitch is "one balance, every API." Instead of an agent juggling a dozen subscriptions and API keys, it draws from a single funded balance and pays per call, with native support for MCP payments and x402. The bet is that agents will want to discover and buy data and services on the fly, and that nobody wants to pre-provision accounts for purchases that have not happened yet. Coinbase's Agentic Wallets play a similar role at the infrastructure layer, giving an agent its own wallet with spending caps, per-transaction limits, and gasless settlement on Base.
Agent-native banking and identity
Skyfire builds payment rails anchored to a "Know Your Agent" protocol, the agent-world echo of KYC, so a merchant can verify which agent it is dealing with before money moves. Catena Labs raised 48 million dollars from a16z crypto, Coinbase Ventures, and General Catalyst to build what it calls an AI-native bank: identity, stablecoin accounts, and compliance designed for software customers rather than human ones. This is the part of the stack that answers the trust question, not just the settlement one.
Merchant-side checkout
On the other side of the transaction sit the companies helping businesses accept agent traffic. Nekuda, backed by Visa Ventures and American Express Ventures, gives merchants structured endpoints for catalog search, cart, checkout, and order tracking, so an agent can buy without scraping a human web page. Stripe's Agentic Commerce Suite does this at scale, with early brands including URBN, Etsy, and Coach, bundling discovery, checkout, payments, and fraud into one integration. Crossmint is building comparable infrastructure with Mastercard and Google as partners.
The networks move in
The incumbents are not sitting still. Mastercard launched Agent Pay for Machines, with a roster that includes Catena, Coinbase, Crossmint, Skyfire, and others. Visa and American Express are funding the startups directly. When the card networks start writing both checks and standards, it is a sign the category has stopped being speculative.
What it means for brands and builders
It is tempting to read all of this as "agents can pay now, so we are done." Payment is the easy part to imagine and the last part to happen. Before any wallet opens, an agent has to do three things your old funnel never asked of a buyer. It has to discover you among the options. It has to trust you enough to act. It has to understand your offer and your price in a form it can parse without a human reading your landing page.
The order of operations
Discover, then trust, then price, then transact. The money rails being built today are the final step. They make a sale possible. They do not make you the one chosen. A funded agent that has never heard of you, or cannot verify you, simply spends its balance somewhere else.
This is the same lesson every interface shift has taught. In the search era, being in the index came before any click. In the generative era, being recommendable by the model came before any mention. We wrote about that transition in why SEO is no longer enough. Agent commerce adds a fourth surface on top, and it inherits the same rule: the infrastructure that lets a buyer pay you is necessary, but it is never what makes them pick you.
Where this is heading
The money rails are arriving faster than most teams expect. Within a couple of years, "can an agent pay us" will be a solved, boring question answered by a checkbox in a dashboard. The interesting question, the one without a settled answer yet, is how an agent decides who to buy from in the first place. That decision happens upstream of any protocol, in the layer where discovery and trust are won or lost.
That upstream layer is the part we spend our time on at Presenc. If you want the strategic version of this argument rather than the field guide, the thesis is here. Either way, the rails are no longer the bottleneck. Being the brand the agent already trusts is.
