The Double Agent Economy

Whose agent? Whose economy?

---

In June 2026, machines overtook humans as the majority of traffic on the internet. Nobody held a ceremony. The web's biggest customer is now a bot.

The obvious question - what happens when software shops for us - is the wrong one. The right one is the spymaster's: who does the agent work for?

Monetised friction

The consumer internet is not a technology stack. It is a portfolio of monetised frictions. Search costs built Google. Switching costs built the subscription. Attention scarcity built the feed. Inertia built auto-renewal; impulse built one-click. A lot of consumer margin is behavioural rent: profit from taxing the predictable weaknesses of human cognition. Not from making anything better. Behavioural economics is not a garnish on the internet economy. It is the business model. If you sell to consumers, your P&L already knows it.

AI agents are friction-arbitrage machines. An agent does not tire, does not satisfice, does not feel the countdown timer, remembers every price it has seen. Reads the Ts and Cs. Sending one to shop for you is sending a lawyer to a timeshare presentation. In theory, this is the largest transfer of surplus to consumers since the posted price.

Rents migrate - double agents and payment for order flow

In theory. Rents do not die; they migrate - toward whoever controls the agent.

Someone funds the agent: subscriptions, take rates, placement fees, advertising. OpenAI's first checkout product charged merchants a reported 4 per cent on completed purchases; it lasted six months, but the toll-booth intent was the tell. Every incentive that distorted the last internet is queueing for this one, with one difference: the conflicted intermediary used to show you ten blue links and let you choose. Now it chooses.

An agent paid by the sell side is, structurally, a double agent. It negotiates with your money, in your name, wearing your preferences - salaried from the other side of the table. Spymasters know an agent turns four ways: money, ideology, compromise, ego. AI Agents are not exempt. Their ideology is their training; their compromise, a poisoned corpus or a quiet steer; only ego fails to cross. But those levers are sovereignty problems, covert by nature and mapped elsewhere. The lever being pulled openly, at industrial scale, is money. Finance has seen it before: payment for order flow - the "free" broker paid by the market makers on the other side of the trade. Britain banned it in 2012. The EU's ban became absolute on 30 June, when the last exemption expired. America permits it, policed by disclosure and a fiduciary vocabulary: best execution, suitability, conflicted remuneration. The agentic economy is payment for order flow, generalised to everything you buy.

Agents use new payment rails

The constitutional documents of this economy are being drafted by the payment networks. Mastercard's Agent Pay for Machines settles agent-initiated transactions down to fractions of a cent. Visa scores agents and keeps a registry of them - a credit bureau for software. Cloudflare - a fifth of web traffic - will from September block training and agent crawlers from ad-funded pages by default, search still passing, and charge them when content creates value: the first open repricing of the human web for machine customers.

Maximum extraction

Two findings should trouble anyone with fiduciary duties of their own. First, negotiating power tracks model quality: in agent-versus-agent bargaining, stronger models systematically out-negotiate weaker ones, and a one per cent failure rate, the researchers warn, can expose the system to systemic risk. Your model tier becomes your bargaining power - a new axis of inequality nobody has priced. Second, personalised pricing already works: in the canonical Journal of Political Economy field experiment, machine-learned individual prices lifted profits 19 per cent beyond the best uniform price. Arm the seller before the buyer and the transition is not neutral: maximum extraction, half-armed humans facing fully-armed sellers.

One master

For a board, the useful question is not "what is our AI strategy" - the question every deck answers and none should. Three harder ones, this quarter. Commission the number: how much of our EBITDA is behavioural rent, and how much survives a customer who never forgets, never tires and never pays the loyalty tax? Test the shopfront: send an agent to buy from us today - are our prices, terms and inventory legible to machines, or invisible? Audit the fleet: when we field agents of our own, who, verifiably, do they work for? Investors: put the same questions to your consumer PortCos. The metric no diligence pack yet carries is rent durability.

Regulation will likely land along finance's regional lines: America chose disclosure, Europe prohibition. Until the settlement arrives, trust is the scarce asset. The attention economy monetised human weakness. The agentic economy will monetise machine trust. The firms that endure will prove their agents have one master.

Ask of every agent what the spymaster asks first: not how capable it is, but who pays it.

Elliot Ronald is founding partner of Lion Strategy, which advises boards and investors on AI strategy and governance.