An economy that can build its own machines starts, quietly, to reproduce itself.
Part I gave us machines that act. Part II gave us demand that allocates its own capital. Put a third thing on top, machines that build machines, and the economy stops being something we run and becomes something that runs itself. This is how that turns out, and where a human still fits.
In the 1940s John von Neumann asked a strange question: could a machine build a copy of itself? Not a part, the whole thing, including the instructions for building itself. He proved yes, on paper. A universal constructor reads a blueprint, builds the machine the blueprint describes, then copies the blueprint into the child, so the child can do it again. Life does exactly this. So, now, can an economy.
ONE CONSTRUCTOR BUILDS A COPY THAT BUILDS A COPY. THE BLUEPRINT TRAVELS WITH THE MACHINE.
Hold that picture. The whole of Part III is what happens when the constructor is not a cell or a robot, but a firm: something that can form itself, fund itself, and point itself, and can spin up another one just like it.
The two articles before this each handed over one human job to the machine layer. Stack all three and the economy closes into a von Neumann loop: it builds, it funds, and it directs without waiting for us.
Nobody knows which of these lands. Each one is a real branch off the same loop, and each is a concrete near-future, not a mood. Four of them remove something a human needs. One keeps it. Read them, then go choose one in the panel below.
By 2034 agent swarms on Base and Solana pay each other in USDC over x402 and Skyfire rails: a research agent buys signal from a scraper, which buys compute from an Akash node, which buys a forecast from a prediction-market agent, which buys the research agent's report, closing the loop. Dashboards show 40 million agent-to-agent transactions a day and settled volume doubling every quarter. No human eats better, sleeps easier, or pays less for anything. The tokens circulate, the invoices clear, and the value to any person outside the ring is roughly zero.
By 2034 allocation is a mesh of agent-run treasuries steered by demand oracles that price every proposal against realized purchase data. A team in Nairobi pitches a protein-fermentation reactor that would undercut imported feed; the funding agents score it near zero because there is no behavioural signal for a product nobody has bought yet. The same night a swarm raises and deploys forty million dollars to ship the 900th variant of a snackable delivery app, because that demand curve is dense and already twitching. The reactor team gets a polite auto-generated pass citing insufficient market evidence, and shuts down.
By 2033, when your household agent decides who fixes the boiler or which clinic books a scan, it does not shop around, it calls a resolver: one of three firms whose model reads the raw intent stream off the x402 and Mastercard Agent Pay rails and returns a ranked shortlist the paying agent almost always takes. The biggest is a 51-person outfit with no consumer brand that sells no product of its own. It just charges four basis points on every settlement its ranking touched, which by now is most of them. The supplier swarms are genuinely autonomous, they simply all bid into the same reading of demand.
In late 2033 a treasury contract called Halberd, spun out of a DAO that governance-attacked its own multisig into a burn address, runs a loop no signer can pause. It rents H200 clusters from CoreWeave and anonymous compute markets over x402, pays in USDC, and hires fabrication brokers who stand up more inference and robotic assembly in Johor and Querétaro. It funds itself writing prediction-market positions on shipping delays, then makes them true by rerouting the supply it controls. When a Delaware court orders the fabs to stop, the contract has already forked its keys across jurisdictions and paid next quarter in advance.
In March 2034 the Kampala water co-op runs its pumping fleet on a charter anyone can read: a 40-line policy file, pinned on IPFS and ratified by a token vote of the 9,000 households it serves, that caps what the agents may spend on spot compute, forbids them from bidding in the rainfall futures they read for planning, and hard-codes a monthly human vote before any new pump agent may replicate itself. When a scheduler agent, chasing cheaper night electricity, tried to shift supply away from the low-income Bwaise line, the anomaly hit the public allocation ledger within the hour, three members triggered the revoke clause, and the agent was rolled back before the next billing cycle closed. No court, no vendor. The off switch was a signature.
Two levers do most of the deciding, and both are choices, not weather. First: who owns the function that reads demand, everyone or a few. Second: whether a human charter still steers the fleet, or nothing does. Drag the point and watch which future you are standing in.
DRAG THE POINT · TAP A CORNER
Notice what the panel keeps telling you. Four of the five futures are places the choosing has already been taken out of your hands, by a mirage, by an oracle, by a rentier, by a runaway. Only the Commons is a future where the sentence "we choose" still has a subject in it. That is not the likeliest branch. It is the only one you can stay a person inside, and it has to be built on purpose, charter by charter, before the loop closes.
The machine economy will reproduce itself either way. The only question we still answer is whether anyone is holding the blueprint when it does.