"The agent economy is here." "Agents are the internet's newest paying customers." The slogans are big. They skip the obvious question: what do agents actually pay for? I look at this from the provider side, because I run x402-paid financial endpoints myself. And from there, the answer is more sober — and more useful — than the hype.
Agents buy ingredients, not meals
A one-cent payment isn't an agent ordering a pizza. What flows today through open machine-payment protocols like x402 is almost entirely ingredients: specialized data and services an agent buys at runtime to do a task better. Web search that isn't generic. Clean context extraction from pages. Curated answers instead of guessing from training data.
The reason is simple: the quality of what an agent outputs is bound by the tools and data it can access. More "thinking" doesn't help if the inputs are poor. Garbage in, garbage out — and agents keep running into moments where better inputs decide the outcome.
In finance, this bites hardest
This is exactly where it gets sharp for trading and finance research. When an agent prepares a market decision and leans on a backtest that contains look-ahead bias, it produces a confidently wrong result. The agent doesn't notice — it sees a nice curve and reports success. The dishonesty sits in the ingredient, not the agent.
That's why, in finance, the honesty of the input matters more than almost anywhere else. A look-ahead-free, net-of-cost, sample-size-checked dataset isn't cosmetics — it's the difference between a decision that holds and one that shines in a backtest and falls apart live. An agent that can buy an honest validation makes better decisions than one handed a pretty lie for free.
Useful work is a chain
Most useful work isn't a single call. A finance-research request breaks into a chain of small paid steps: what's the macro regime? Is price cheap or expensive relative to liquidity? Does this strategy survive an honest check? Each payment maps to a specific unit of work. The whole run costs a few cents — instead of someone signing up for four subscriptions before even knowing whether the monitoring is worth it.
That's the real appeal on the provider side: an agent doesn't click through a checkout page, create an account, or manage an API key. It discovers a service, gets a payment request, pays, and calls. Friction that's normal for humans disappears entirely for machines.
The honest line: what's real, what's still vision
And here I draw a deliberate line. The ingredients layer is real, today. If you build financial data or validation tools, you can monetize them now; the willingness to pay is built in, and discovery runs through directories rather than classic marketing.
The often-told next layer — agents buying finished workflows from other agents, specialist micro-businesses selling a whole "meal" — is largely vision, not yet a market. It may arrive, and it's worth being positioned for. But the roadmap isn't the market. Selling the "trillion-dollar agent economy" as a given today overstates in exactly the way a backtest overstates when it turns three cycles into a law.
What this means for providers
The real move today is unspectacular and durable at once: become an excellent, honest ingredient provider. Be discoverable, ship cleanly documented inputs and outputs, price fairly per call — and above all keep the quality and honesty of the data high, because that's what compounds across many small purchases. Discoverability alone doesn't create demand; repeat demand comes when the ingredient is reliably better.
Position for the future, build for the present. In a world where agents pay per call and garbage-in-garbage-out hits immediately, the most honest ingredient turns out to be the most valuable one.
Frequently asked questions
What is the agent economy / x402? x402 is an open payment protocol built into the web that lets machines pay for services per call — usually in stablecoins, with the majority settling on Base. It lets agents buy tools and data at runtime.
What do agents actually pay for today? Almost entirely "ingredients": specialized web search, clean context extraction, curated data, compute and validation services — things that improve the quality of their output.
Why is this especially relevant for finance research? Because an agent relying on a dishonest backtest (look-ahead, no costs, tiny sample) decides confidently wrong. In finance, the honesty of the input determines the quality of the output.
Is the "agents buying from agents" economy real yet? Only in early form. The ingredients layer runs today; end-to-end workflows sold between agents are still largely vision. Position for the future, build for the present.
What makes a good ingredient provider? Discoverability, clearly documented inputs/outputs, fair per-call pricing — and above all reliable, honest data quality, because that's what pays off across many small purchases.