Hi,

This is Plain Strata, the Tuesday Pulse.

For as long as software has existed, it could not really pay for anything. The whole payment system is built around a person: a name on a card, a bank that trusts that name, a fraud model that assumes someone is sitting at the keyboard. Take the person away and commerce stops.

This year that changed, and it changed twice at once. Two rival groups shipped two different ways for software to pay on its own, and both built their answer on the same forgotten corner of the web: a status code reserved in the mid-1990s and left empty for twenty-nine years.

Here is the strange part. The same handful of companies that move the world's money, Stripe, Visa, Mastercard, are not backing one answer over the other. They are building or funding both, at the same time.

When the biggest players in a market refuse to pick a side, that tells you something about the market. Here is what it tells you here.

Listen:

Spotify: [SPOTIFY LINK]

Apple Podcasts: [APPLE PODCASTS LINK]

YouTube: [YOUTUBE LINK]

The full piece, no need to click through:

Somewhere this week, a piece of software rented five minutes of a paid browser session, ran its errand, and let the session expire. Somewhere else, a piece of software paid to have a letter printed on real paper and mailed to a real address. Somewhere else, a piece of software ordered a sandwich. No person clicked buy in any of these. No person even watched. An agent decided it needed something, found the price, paid it, and moved on.

That is not a demo. It is commerce running this week on two separate payment rails that both went live in the last four months, both built for exactly this: software that spends money on its own. Each transaction is small, often a few cents, sometimes less than a cent. None of them would be worth a human's time to approve one at a time, which is exactly the point: this only works if no human has to.

Here is the story. Software has never been able to pay for things by itself, not really, because the entire payment system is built around a human: a name on a card, a bank that trusts that name, a fraud model that assumes a person is at the keyboard. This year that changed. And it changed twice, independently, using the same unused piece of the internet, which is why the companies that move the world's money are now hedging their bets on two different answers to the same question at once.

An agent, in the sense this show uses the word, is software that can decide something and act on it without a person approving each step. Until this year, the moment an agent needed to buy anything, real commerce stopped and a human had to step back in: enter a card number, approve a charge, accept a subscription. That single choke point is what "agent payments" means solving. Not a new kind of money. A new kind of yes, one a machine can say on its own, safely enough that a company will actually ship it.

The reason this took until 2026 is not that anyone lacked the idea. It is that every existing payment rail assumes a person: a name a bank can call, a signature a court can enforce, a fraud model trained on how humans behave when their card is stolen. An agent has none of that. It has no name a bank recognizes, and it can act thousands of times a second, which looks exactly like the pattern a fraud system is built to block. Solving agent payments meant building something a bank did not have to trust as a person at all.

The two answers that shipped this year both start in the same strange place: a status code.

Every time a browser or an app talks to a website, the website replies with a three-digit number that means something specific. 200 means here is your page. 404 means nothing is here. In the mid-1990s, when the first official rules for how the web talks (HTTP) were being written, the people writing them set aside the number 402 and labeled it "Payment Required." The idea was that someday, paying for something online would be as automatic as loading a page: hit a paywall, get a 402, pay, get your content.

Nobody built that part. For twenty-nine years, 402 sat in the HTTP rulebook completely unused, the most famous feature the web never shipped. Every browser and server has known about it the whole time. Nothing ever answered it.

This year, two separate groups looked at that empty room in the rulebook and decided to move in.

The first is x402, built by Coinbase and given, on April 2, 2026, to a new nonprofit called the x402 Foundation, run jointly with the Linux Foundation, the same organization that stewards the Linux operating system. (The launch date, 4/02, is a small joke pointed at the status code itself.) x402 answers a 402 in one shot: an agent hits a paywall, gets the 402 back with a price attached, pays it instantly in a stablecoin, a dollar-pegged digital token, and gets what it asked for. One request, one payment, done. By this summer, Coinbase's CEO was citing more than 160 million of these machine-to-machine payments processed since launch.

The second is the Machine Payments Protocol, or MPP, built by Stripe together with Paradigm and Paradigm's own payments-focused blockchain, Tempo. It went live on Tempo's mainnet on March 18, 2026, over two weeks before x402 had its formal foundation. MPP answers the same 402 differently: instead of one payment per request, it opens a session, a running tab an agent can draw against for a while, and it settles that tab in whatever an agent has on hand, a stablecoin, a card, even the Lightning Network. That is what let an agent rent a browser session by the minute or run up a small tab ordering a printed letter or a sandwich: those are sessions, not single purchases. Within its first week, MPP was reportedly wired into more than fifty services, including OpenAI, Anthropic, and Google's Gemini.

Here is the detail that turns this from two products into one story worth stopping on: Stripe is a founding member of the x402 Foundation. Visa and Mastercard are too. The same handful of companies that are building x402 are, at the same time, building or plugging into its rival. Visa has confirmed support for both rails. Nobody in this story is betting on one horse.

There is a name for what the internet's engineers were originally trying to build with a single empty status code like 402, and it is worth naming in full because it explains why this week's split matters.

The internet is shaped like an hourglass. At the top, a huge and constantly changing variety of things people build: apps, websites, games. At the bottom, an equally huge variety of physical wires and radios: fiber, satellite, your phone's antenna. In the middle, at the hourglass's narrow waist, sits one single, boring, unglamorous rule that everything above and below has to pass through: the Internet Protocol, or IP. Because everyone above and below agrees to speak IP, and only IP, anything built on top of it can reach anything built on top of anything else. That narrow waist is the whole reason the internet works as one internet instead of a thousand incompatible ones.

Payments were supposed to get their own narrow waist. 402 was reserved to be exactly that: one universal, boring "pay here" moment that any agent, on any service, could answer the same way. What shipped this year is two waists, competing for the same slot in the hourglass, at the same time, backed by overlapping companies.

The honest implication is that a universal layer only works if it stays singular. The whole value of a thin waist is that a builder plugs into it once and reaches everyone. Two candidate waists means a developer building an agent that needs to pay across the open web has to speak both languages, or bet on one and lose access to whatever only accepts the other. That is exactly the fragmentation a shared protocol exists to prevent, showing up anyway, this time in money instead of network cables.

The honest open question is whether this settles the way it should. Either one rail wins outright and the other fades, or a translator layer forms on top, something that lets an agent speak one language while it actually gets converted to whichever rail the other side wants. Visa quietly supporting both is the first real evidence of which way this goes: not a bet that one wins, but the beginning of exactly the kind of bridging layer that a working thin waist was supposed to make unnecessary. Whether that bridge becomes the field's real third protocol, sitting above the first two, is the thing worth watching.

There is a version of this that has happened before. Video once had two competing tape formats, then a handful of competing disc formats, before the format question stopped mattering because everything moved to streaming, a layer above the format entirely. Money for agents may be heading somewhere similar: not a winner between x402 and MPP, but a point where the choice between them becomes invisible to whoever is building the agent, handled by something else underneath. That would be a third rail in truth, not a second one wearing a different badge.

Watch whether Visa's dual support for x402 and MPP stays a hedge or turns into a real bridge that lets an agent built for one rail transact on the other without knowing the difference. And watch the adoption numbers themselves, x402's raw payment count against MPP's service integrations, since whichever curve bends harder in the next few months is the nearest thing this story has to a scoreboard.

The two voices are AI. The research and writing are mine.

Decentralized AI, layer by layer.

Dastan