Every market that has ever existed shared one property so universal that nobody thought to name it: somewhere in every transaction, there was at least one human present. A buyer, a seller, a broker, a clerk. Markets were places where people met, even when the meeting was a phone line or a form. The whole apparatus of commerce, contract law, receipts, trust, fraud, negotiation, was built as a description of what happens between people when value changes hands.
The digital city is now running the first transactions in economic history where that property fails. An agent, dispatched on a human's errand, encounters a resource with a price. The price was set by software, according to a policy. The agent evaluates it against a budget, pays from a wallet it carries, receives the goods, and files a receipt. Somewhere upstream a human wanted something, and somewhere downstream a human will get paid, but the transaction itself, the offer, the evaluation, the payment, the settlement, occurred between two pieces of software in a fraction of a second, witnessed by nothing but a log.
This is not speculative plumbing. The payment protocol for it processed 169 million transactions in its first year, runs under the Linux Foundation with the largest infrastructure companies as members, and settles in stablecoins in under a second at costs below a cent. The company that carries a fifth of the web's traffic built the tollgate into its edge and described the future plainly: soon the agents "will carry wallets and buy what they need without a person in the loop." Sites behind that network already issue over a billion payment demands a day to machine visitors. The pipes are laid. What flows through them is just beginning.
Think about what a mature version looks like, because the shape is visible from here. An orchestrator agent gets a human errand: plan the offsite, source the part, prepare the filing. It decomposes the errand and hires. A data agent, paid per record. A translation agent, paid per document. A verification agent, paid to check the others. Specialist agents selling services to generalist agents, prices discovered per request, providers swapped mid-task for a tenth of a cent's advantage. An economy of subcontractors, deep and fast and wholly submerged, where the entire supply chain between the want and the outcome consists of parties that cannot want anything.
Now notice what goes strange in that picture. Markets, as we built them, ran on friction that turned out to be load-bearing. Negotiation took time, and the time was where judgment lived. Reputation accumulated over years, and the years were the proof. A price that changed hourly was a scandal; the stickiness was a courtesy to human attention. Machine markets have none of these frictions, which sounds like efficiency until you remember what we learned the last time we removed humans from a market's tempo. High-frequency trading was the pilot program for commerce without witnesses: prices discovering themselves in microseconds, strategies interacting in ways no participant designed, and occasionally, famously, the whole thing convulsing in a flash crash that took regulators months to reconstruct from the only witness available, the log. That is not an argument that agent commerce fails. Equity markets work. It is an argument about what supervising such a market is like: you do not watch it, because it cannot be watched. You audit it, afterward, at human speed, hoping the logs are honest.
Which makes the log the most important civic document in the new economy. When neither party to a transaction can be deposed, testify, or remember, the record is not evidence of the event; it is the event's only residue. Every question we used to resolve by asking people, what was agreed, who knew what, was the price fair, becomes a question of provenance: does the trail exist, is it complete, can it be trusted. Commerce without witnesses can be honest commerce, but only if the receipts are constitutionally incapable of lying, and building receipts like that is now infrastructure work as fundamental as the payments themselves.
And here is the human experience of all of it, the part the legacy residents should actually prepare for: the economy gets quieter. Not smaller, quieter. A growing share of buying and selling simply stops being observable from where we stand. The invoice arrives aggregated: your agent spent four dollars this week, across eleven hundred transactions, on your behalf, optimally. Every line item is available and no one will read them, the way no one reads the itemized calls on a phone bill. We will live inside a roaring marketplace and hear almost nothing, and the monthly summary will be the porthole through which an entire economy is glimpsed.
The old markets had a saying that priced their honesty: caveat emptor, let the buyer beware. The buyer, now, is software, and it is never beware and never charmed; it is exactly as careful as its instructions. So the burden moves one link up the chain, to the writing of the instructions and the auditing of the trail, which is to say it moves back to us, in the only two places we still appear in the transaction: before it, and after.
And the goods themselves are next, because none of this architecture is native to software; it is native to errands, and the oldest errands are physical. Watch industrial procurement, my corner of the world, because it will be the proving ground. A buying agent dispatched to source a machined part faces the problem every market solves first: discovering who can sell. It will not tour factories, notice the clean aisles, or work a trade-show floor. It will query for capacity the way it queries for anything, which means the industrial economy is about to learn what retail learned a decade ago: a seller a machine cannot read is a seller that does not exist. So the registries come first, capacity stated in structured form, capabilities, equipment, materials, tolerances, certifications, the market square rebuilt as a schema, and whoever keeps that registry keeps the introductions for an entire industry.
Then comes the harder problem, the one that makes physical goods the real test of witness-less markets: trust. A four-dollar data purchase can afford to be wrong. A machined part cannot. Between humans, a quote is believed because of everything around it, the relationship, the reputation, the plant tour, the handshake. Between machines there is nothing around it. The trust has to travel inside the artifact, which means the winning quote in a machine market is not a number but a derivation: a price with its physics attached, the formula, the constants, the limit the cut was scored against, checkable by the counterparty's own software in the second it arrives. Reputation took years because humans verify slowly. Derivation takes milliseconds. In a market without witnesses, the shown work is the witness.
Assemble the pieces, the registry that makes sellers legible, the quote that carries its own proof, the rails that settle machine to machine, and a part gets specified, quoted, purchased, and built with the first human witness being whoever opens the box, holding the one thing no market has ever traded away: the decision that what arrived is right.
We used to be the market. Now we are its opening bid and its closing audit. Everything in between has stopped waiting for us.