The agentic economy was never an intelligence problem. It was a settlement problem.
For most of the 2020s, the story was about cognition: agents that could reason, plan, and negotiate. That was the interesting part to watch and the wrong part to bet on. An agent that drafts a contract at three in the morning is a parlor trick if it then waits until a compliance officer starts work at nine to move the money. The capability that reorganized commerce was duller — the moment agents got rails that settled at their speed instead of ours. From 2036, that is the line the whole decade pivots on.
What everyone predicted
The consensus forecast got the magnitude roughly right and the mechanism almost entirely wrong.
The projections were enormous and, it turns out, defensible. Goldman Sachs estimated in 2023 that generative AI could lift global GDP by about 7 percent — nearly $7 trillion — and raise productivity growth by roughly 1.5 percentage points over a decade of broad adoption.1 McKinsey put its annual value at $2.6 to $4.4 trillion, with banking, high tech, and life sciences landing hardest as a share of revenue.2 Those numbers held up.
What the consensus missed was where the bottleneck sat. The 2024 mental model was a smarter worker: an agent doing knowledge work faster. The real shift was structural. Agents did not just need to think; they needed to transact — to hold value, pay a counterparty, and have the payment be final without a human in the loop. The question was never how clever the agent was. It was who controlled the layer where its decisions became money.
Force one: settlement at machine speed
Run a fast agent through legacy financial plumbing and you get a queue. Correspondent banking, clearing windows, wire cutoffs, business hours — each is a place where a machine waits on a human institution. The agentic economy's first enabling force was the removal of those waits.
The fix arrived as standards, not as one product. In September 2025 Google published the Agent Payments Protocol, an open framework built with more than sixty firms — Mastercard, PayPal, American Express, Coinbase among them — to let agents pay and be paid with cryptographically signed authorization.3 Coinbase's x402 extension wired stablecoin settlement directly underneath it, so an agent could pay in USDC with one signature and get instant, final settlement.4 Visa opened its Intelligent Commerce program to let agents transact on its network, and introduced a Trusted Agent Protocol to tell legitimate agents apart from bots at checkout.5 The institutional rails moved in parallel: JPMorgan's blockchain unit, Kinexys, had processed more than $1.5 trillion in notional value and was clearing over $2 billion a day, around the clock, across five continents.6
This is the enhancement — the law of media that says every technology intensifies some capability. Here it intensifies velocity of settlement. A contract that meets its conditions executes; value moves; the transaction is final. No callback, no cutoff, no morning.
Force two: trust as a data structure
The second force is what made machine-speed settlement safe enough to use. In the old economy, trust was a social technology — relationships, references, years of dinners. Two agents from competing firms have none of that. So trust got rebuilt as infrastructure: a counterparty agent's identity, payment history, and authorization scope, all cryptographically verifiable before anyone agrees to anything.
This is the tetrad's retrieval — the law that says new media bring back something old in new form. What returns is the oldest commercial instrument there is: reputation as collateral. The medieval merchant's standing in the guild, the correspondent bank's know-your-counterparty, the credit reference — all of it comes back, now as a signed record that updates in real time and cannot be quietly edited. The agent with a clean, verifiable history closes the deal. The anonymous one does not.
And here is the obsolescence, the law of what gets pushed aside. The casualty was not the human negotiator first; it was the intermediating apparatus — the clearing house as a place, the settlement delay as a fact of life, the nine-to-five compliance window as a hard constraint. The Bank for International Settlements made the direction unmistakable in 2025, when its researchers showed a general-purpose AI agent could perform intraday liquidity management in a wholesale payment system — maintaining buffers, prioritizing payments under stress — without domain-specific training.7 When the treasury desk's core judgment can be replicated by a model with no special schooling, the desk is no longer load-bearing.
The reversal
Every enabling force in this story was sold as decentralization. The rails were open. The reputation ledgers were public. The promise was that economic power would distribute — that any firm with a competent agent and a clean record could transact with any other, no gatekeeper required.
Pushed to its limit, it flipped. This is the law that matters most and gets skipped most often: what a technology reverses into when you run it all the way out.
The best agents came from the firms with the deepest data and the largest compute budgets. The richest reputation histories accrued to whoever was early and large. The "open" rails consolidated around a short list of protocol stewards and settlement providers — the same payment networks and money-center banks that ran the old system, now operating the new one's plumbing. Machinery built to remove gatekeepers manufactured a smaller, more powerful set of them, one layer down where customers never look.
The bill came in two other forms the optimists under-priced. Speed cut both ways: a compromised agent does not steal slowly, and crypto theft still ran to about $2.2 billion in 2024 alone, with private-key compromise the single largest vector.8 And accountability lagged by years. When an autonomous agent enters a contract that goes wrong, who owns the loss? The EU's AI Act began assigning that liability — chiefly to the providers and deployers of high-risk systems — but most jurisdictions had no answer, and the agents were already transacting in the gap.9 The tool built for inclusion became an instrument of concentration, surveillance, and unowned risk. That is the reversal, and it is the part the 2024 forecasts left out.
The lesson
The capital followed the rails, not the cognition. Venture funding into agentic-AI startups reached roughly $2.8 billion in the first half of 2025 alone, and the durable winners were rarely the cleverest models.10 They were the firms that owned a layer of the stack: identity, reputation, settlement.
That is the principle worth keeping. In an economy of autonomous agents, intelligence is a commodity and the settlement layer is the franchise. Whoever controls where a machine's decision turns into money controls the economy that machine operates in. The firms that understood this stopped asking whether to deploy agents years ago. They asked which layer to own — and they were right to, because the answer set the terms for everyone who arrived later.
The window to shape that layer deliberately — its identity standards, its liability rules, its concentration limits — was open in the mid-2020s and closed the way these windows always close: quietly, by default. The economy that never sleeps was built while almost everyone was asking the wrong question. Look back from 2036 and the rule carries forward plainly: decentralization is a property of the layer you can see; centralization hides one layer down, in the plumbing nobody audits until it owns them.
This piece is speculative journalism written from a fixed 2036 vantage. The protocols, institutions, figures, and research cited below are real and sourced as of 2025; the 2036 end-state and the dated figure captions are one plausible projection, not a forecast of an inevitable outcome. The reader is meant to know which is which.