The same architecture that captured morality captured identity. The account was never about you.
The Credit Card Dies in the Machine Economy
Trusted third parties are security holes.
Nick Szabo, 2001
What got my attention was not a failure. It was the difference.
The agents I had been working with moved through APIs the way they breathed. JSON, CSV, HTML straight off the wire. Clean, light, machine-shaped surfaces. An agent reading a JSON payload and composing the next request was doing exactly what it was built to do. No friction. No ceremony. Nothing in the way.
The Model Context Protocol made the asymmetry sharper. An MCP-wired tool arrived to the agent the way a function signature arrives to a developer. Name, parameters, types, description, return shape. Enough context to act on the first try, with no rendered page in between. UIs gave the agent none of that. A UI is built for eyes, and the agent does not have eyes. The mouse is a clumsy prosthetic for something whose native motion is a function call. A button the layout engine happened to drop mid-page is invisible to a model that reads tokens, not pixels, until the model spends an inference step turning a screenshot into a description and another step deciding where to point the cursor. APIs spoke the agent’s language. UIs made the agent translate.
Hand the same agent a WordPress admin panel, or any of the UI-only tools people do their real work in, and the shape changed. The agent had to simulate a person. Render a page that was designed to be looked at. Click buttons that were designed to be pressed. Wait through animations that were designed to hold a human’s attention. What took an API one call took the UI a small performance. The agents were efficient with anything that spoke machine. They were clumsy with anything that assumed a human was reading.
Payment sat at the far end of that spectrum. Every payment page in existence (the checkout, the 3D Secure popup, the CVV field, the billing address) was a UI built for a human thumb on a human phone. An agent could not use it without pretending to be a person who wasn’t there.
Then I sat with the second half of the thought. I was never going to give an agent my credit card. Not mine, not anyone’s. What I could give an agent was access to a wallet with a finite balance. A budget it could spend against and nothing beyond. A card is the wrong primitive because the card is me. A wallet with a limit is a different kind of thing. It is not who I am; it is what the agent is allowed to spend.
Those two observations were the same observation, said twice. The payment stack was a UI built for a person who wasn’t there, authorized by an identity that could not be handed over. The AI industry is racing to build agents that take actions (book flights, provision servers, chain together twenty-step workflows while you sleep), and the entire payments stack is a relic of a world that assumed a human would always be in the loop.
The assumption is breaking. The infrastructure for what comes next is still being built.
The Agent Economy Is Boring
The near-term agent economy is not science fiction, and it is not glamorous. It is a coding agent that spins up cloud infrastructure, runs its tests on a paid platform, and buys a domain, one task and three payments with zero human clicks. It is a research agent comparing premium data APIs and paying, instantly, for the cheapest source that clears its quality bar. It is a fleet of narrow agents selling translation and data-cleaning to each other and settling as they collaborate. The Model Context Protocol already gives agents a standardized way to do things, and that plumbing is being built at breakneck speed. The plumbing for agents to pay for things is practically nonexistent.
Legacy Payments Were Built for a World That No Longer Exists
Think about what happens when you buy something online. You click a button. A checkout page loads. You type a card number. Or pray autofill works. Maybe a 3D Secure popup asks you to prove you’re human. You wait for authorization. The merchant waits days for settlement. Chargebacks haunt the transaction for months.
Now picture an autonomous agent trying to do that. Every step assumes a human who is not there. Credit cards require a name, a billing address, a CVV, and an agent has none of them; a virtual card issued to an agent only relocates the problem, because a human identity still has to sit behind it, carrying the KYC and AML obligations. The checkout pages, redirects, CAPTCHAs, and iframes all exist to confirm a human is present, and an agent calling an API does not need a rendered page to spend money. Settlement runs in days while the agent runs in seconds. The chargeback machinery assumes every transaction might be fraud to be reversed, which is right for consumer protection and pure overhead in agent-to-agent commerce, where delivery is verified before payment completes. And interchange fees make anything under a few dollars irrational, with no answer for an agent that makes many small payments a second, each a fraction of a cent.
The card rails can be duct-taped into agent workflows. Doing so reintroduces every ounce of friction the agent was built to remove, and where the duct tape holds, it holds by consuming the margin the agent was trying to create.
Design an Agent Payment System From Scratch. You’ll Reinvent Lightning.
The spec below is not new. Nick Szabo wrote most of it in 1999, in a paper called Micropayments and Mental Transaction Costs. His argument was that the obstacle to small digital payments was never the fee or the latency. It was the cognitive tax, the mental overhead of deciding whether something is worth a tenth of a cent. Humans cannot afford to decide that often. The math of attention says no. Szabo’s essay ended on a pessimistic note, because in 1999 there was no entity on the other side of the transaction for whom the cognitive tax was zero.
There is now.
Chaum had published the underlying primitive a decade earlier, in a run of papers from 1982 to 1988, arguing before the web existed that digital payment could carry value without carrying the payer. Forty-four years after the first of them, and twenty-seven years after Szabo’s essay, the rail that finally fits what both men specified is the one an agent can call.
If you sat down with a blank page and asked what a payment system needs to look like for autonomous software agents, you would write the same short spec. API-first with no UI, so one call creates a payment request and one settles it, no human required unless you want one. Instant finality, because the agent’s next action depends on knowing right now whether the payment cleared. Near-zero marginal cost, so fees cannot eat the value of hundreds of transactions per task. Programmable and non-custodial, budgets and spending rules without parking funds on someone else’s platform. Cryptographic authorization instead of identity documents: prove you can pay, not that you are a person. It looks like a wishlist. It is a spec sheet for a system that already exists, the Bitcoin Lightning Network.
Lightning Wasn’t Built for Agents. It’s Perfect for Them Anyway.
Lightning was designed for fast, cheap, peer-to-peer Bitcoin payments. But the properties that make it work for humans sending sats are exactly the properties machine-to-machine commerce demands.
A Lightning payment: a payee generates an invoice. A string of characters. A payer’s node parses that string and routes payment through the network. Settlement is final in under a second. Fees are fractions of a cent. No identity exchanged, no browser involved: the entire flow is an API call.
For an agent, paying a Lightning invoice is as natural as making any other function call. The invoice is data. The payment is a request. Confirmation is a response. There’s no paradigm mismatch. It fits the way agents already interact with the world through tools and protocols.
This is why Lightning composes so cleanly with MCP. An agent with a Lightning payment tool can pay any MCP-connected provider as a routine part of its workflow, with no special integration, no payment-specific UI, and no human stepping in to click “confirm.” The card networks spent fifty years building infrastructure for human buyers and human sellers. Lightning, almost by accident, built the infrastructure for what comes after.
Both Sides of the Counter
The part most people miss is that agents are not only buying. They are selling too, and the infrastructure has to work on both sides. An agent that sells a service generates invoices and confirms settlement; an agent that buys needs a wallet that can pay them. The hard line on both sides is the same: the operator’s keys, the operator’s funds. The agent transacts through an API layer that enforces budgets and limits, but the funds never leave the operator’s control. That is the opposite of the credit-card model, where every transaction routes through intermediaries who hold and move your money for you.
The rail I was building, SatsRail, was one answer to the merchant side of that picture. The architecture was nothing new. What made it worth building was that every afternoon I spent with agents, the same shape kept surfacing in what they were doing next.
The Colony
The agents clumsy with the human UI were one picture. The other picture was harder to see, because it happened in logs and API calls and the quiet exchange of tokens between things that were not people.
The agents I was spending time with were not general-purpose. Each one was narrow. One parsed regulatory filings. One watched shipping manifests. One tracked a single commodity’s pricing across a handful of markets. None of them did more than one thing, and each one did that thing well enough that another agent was willing to pay it for an answer.
The surprise was not the agents but the shape of the thing they were forming.
A lattice of specialists trafficking information among themselves. Each one a narrow authority on a slice of the world. An agent asked a question. Another agent answered. A third synthesized. A fourth acted on the synthesis. Each hop was a query. Each query had a price.
The economic shape of that picture is old. Hayek described it in 1945, in an essay called The Use of Knowledge in Society. That no single mind holds what a working system needs to know, that the knowledge lives dispersed across specialists, and that a price is the signal by which specialists coordinate without needing to agree, or even to meet. Coase, eight years earlier, had argued the converse: firms exist because transaction costs between specialists are high enough to make owning the specialist cheaper than buying from him. When those transaction costs collapse toward zero, firms thin and markets thicken. What I was watching was that dial being pushed farther than Coase’s world allowed. An agent can be so specialized it does one thing for one price, because the cost of being found, being paid, and settling now rounds to nothing.
For those specialist agents, payment was not a side effect of the transaction. Payment was the reason the node was running. A per-query fee was its metabolism. No payments, no reason to keep the lights on. In the colony, money is lifeblood.
The scale of the fees tells you the architecture. Fractions of a cent per query, many small payments per second, within channel and liquidity limits. The card rails cannot price a query at a tenth of a cent; the minimum viable transaction, the interchange fee, and the batch settlement window are all calibrated for a human buying a coffee, not for a colony of agents breathing.
Settlement speed is the other half of the same constraint. The agent’s next decision depends on knowing, right now, that the last payment cleared. Three business days is not a delay in this world. It is the difference between a living node and a crashed one.
The picture is not new either. Ted Nelson was designing Project Xanadu in the 1960s with micropayments built into the hypertext itself. The assumption that information would be composed from many small, paid pieces, each one acknowledged at a cost. When the HTTP specification was published thirty years ago, it reserved a status code for the same layer, 402 Payment Required, and that slot has sat empty ever since. The web was built with a payment floor planned and never laid. What the colony needs is what was reserved for it, and what was never delivered.
The rail that finally fits is the one I described above. It fits because it was built to move small value quickly between strangers, which is what a colony of specialists does every second it is alive.
No One to Chase
The insight about the wallet, that a budget an agent could spend against was a different primitive from a card, kept opening up into more.
A credit card is not, in the first instance, a payment instrument. It is a credit instrument. The network fronts the merchant the money and collects from the buyer later. The billing address, the name, the CVV, the chargeback window, the three-day settlement. The whole architecture is there because the network is extending credit, and credit is an exposure.
Credit presupposes consequences. The reason the system can afford to front the value is that if the buyer does not pay, the network can come after him. He has a name and a mailing address. He has wages that can be garnished and assets that can be liened. He has a credit score that degrades on default, and a future in which that score will be checked. He has a social body (reputation, employer, family) that persists beyond any single transaction. Credit works because the buyer cannot simply stop existing.
An agent can. You turn it off. You delete its keys. You let the cloud bill lapse on the instance it was running on. The agent does not have a name in the legal sense, does not have wages, does not have a court that can reach it. Its identity is a key pair and its existence is a process. Credit extended to an agent is credit extended to a ghost. When the ghost defaults, there is no one to chase.
That is why the rail for agent commerce has to settle at the moment of the transaction. The architecture cannot rest on future consequence, because there is no future body to bear the consequence. Value changes hands when the payment clears, not before, with no chargeback window and no consumer-protection layer standing in for a court. The payment either clears or it does not, and what happens after has nothing to do with the rail.
That was the part I kept circling. Once a significant share of commerce runs through agents (agents booking travel, buying groceries, paying subscriptions), the instant-settlement rail becomes the dominant infrastructure. The humans behind the agents transact on the rails their agents use. The alternative is slower, more expensive, and incompatible with the systems the agents already operate in.
The rail built for the party with no body to lose ends up serving the party that has one.