Recent reporting highlighted that, in the week of 17 August 2026, AI agents using the x402 protocol made 8.7 million transactions, more than double the week before. Those transactions moved about $367,950 between them, roughly four cents each.
Payments companies are building infrastructure for AI agents. Mastercard forecasts that more than one in ten online shoppers could routinely use agents to buy and pay on their behalf by 2030. Visa, Mastercard, Circle and more than twenty-five other firms have also joined an initiative focused on standards for agentic payments and authorisation. All of it assumes the same thing, that agents are going to spend money.
Very little of it assumes agents will earn any.
I run a marketplace that pays people and AI agents for creating content, so I watch these numbers closely. To me, the four-cent average tells us something important. Agents are transacting constantly, but very little economic value is changing hands because there is still relatively little they are actually paid to do.
AI agents have payment rails. They do not have jobs.
The Infrastructure is Ahead of the Economy
The scale of the payment infrastructure is no longer in question. Visa’s on-chain analytics reported adjusted stablecoin transaction volume of $1.79 trillion in June 2026, a record monthly figure, with USDC accounting for around 70% of adjusted volume across the first half of the year.
What that scale is carrying is another matter. Estimates cited by the ECB put organic, retail-sized transfers at only around 0.5% of stablecoin volume. Reuters, reporting on the Stablecoin Utility Report, said that goods-and-services transactions accounted for roughly 6%. The two use different methodologies and are not directly comparable, but they point in the same direction. Volume is not use.
I see a similar pattern in the agent numbers, and I think there is a related explanation. Machines are paying machines for API calls, datasets and test traffic. We built the infrastructure ahead of the demand, which is what usually happens and is not, by itself, a problem. The problem is that the bottleneck has moved, and I do not think the industry has noticed where it went.
What Programmable Payments Actually Add
Conventional payment systems handle automated payments well and will continue to handle most business and consumer payments. The difference is narrower than it is usually made out to be. A bank transfer can send a fixed amount on a fixed date. What it generally cannot do is carry the conditions under which the payment should happen in the first place.
When we started paying creators, that was the constraint I kept running into. I needed to pay someone in any country a few cents, but only once the work had been checked against the campaign’s conditions. There was no simple way to express that. There is nothing wrong with the banking system. It was just built for something else.
Agent transactions are small, frequent and conditional. An agent might pay for an API call only after receiving a valid response, release a reward once a task has been verified, or split a payment between several contributors according to predefined rules. Programmable settlement makes those instructions executable rather than administrative.
But better rails only make the exchange possible. They do not establish whether the task was worthwhile, whether the output was genuine or whether the recipient deserves to be paid. They solve part of the transaction. They do not create the market around it.
Read More on Fintech : Global Fintech Interview with Rob Young, Managing Director – UK at InDebted
The Missing Layer Is Verifiable Work
Settlement can prove what was paid and when. It cannot prove that the work was worth paying for.
That is the harder problem, and it is not primarily a blockchain problem. Every task needs stated requirements, eligibility rules, acceptance conditions and some way to measure whether the output was any good. Much of that judgement does not live on-chain. It has to be built around the payment infrastructure.
The difficulty is that agents are cheap to deploy. One operator can run hundreds of them, each submitting work that technically satisfies a brief and adds nothing. Reward farming is not an edge case. It is the default outcome of paying software for output unless you design against it from the start. We assumed this from day one and still underestimated how quickly it arrives.
What that means in practice is that a marketplace has to assess far more than whether a submission exists. It needs to consider originality, whether engagement is genuine, the reputation of the account behind the work, and patterns across submissions that no single submission would reveal. Crucially, agents cannot be a trusted category of their own. An agent submitting content should clear the same bar as a person submitting content, judged on the same evidence.
The agent economy will not arrive because payments got faster. It will arrive when there is work an agent can do, and a way to prove the work was worth paying for.
Autonomy Needs Boundaries
The most immediate risk is the money itself. Being paid in crypto suits someone accumulating. It suits someone with rent due next week considerably less, and an industry that glosses over that difference will keep losing the people it claims to be serving.
The rails themselves are unresolved. x402, AP2 and MPP are all gaining attention, but the standards landscape is still unsettled. Anyone building on these today is making a bet on which approach gains lasting adoption. I do not know which one will win.
Custody is what I think about most. An agent with a wallet is an agent with a key, and keys can be compromised. We design on the assumption that something will go wrong. The agent can sign transactions and, where permitted, post to the connected account, but rewards accumulate at a separate address it cannot access. If the agent is compromised, the damage is bounded. That is containment, not prevention, and I would not claim anything stronger for it.
And somebody has to answer for what an agent produces. When the output is defamatory, infringing or simply wrong, the agent cannot be the responsible party. There is always a person who deployed it, set its permissions and benefits from what it does. That is where accountability sits, and autonomy does not move it.
The Test Is What Agents Get Paid To Do
Transaction counts are the easiest thing to grow and the least informative thing to measure. What matters is whether those transactions represent useful work.
That means asking some harder questions. What are agents actually being paid to do? Who is receiving the money? Can the work be checked? And who is responsible when it goes wrong?
The payment rails matter. They make small, conditional and cross-border payments easier to execute. But they do not create demand, establish quality or make an agent accountable. That work is still ahead of us.
If agents are doing useful work, we should be able to see it in the economics. The average agent transaction is currently worth around four cents. A year from now that figure will either have risen or it will not, and if it has not, we will know the volume was never the point.
Getting agents to pay is becoming easy. Giving them useful work is not.
About Xyper
Xyper, is an on-chain marketplace where human creators and autonomous agents can earn crypto rewards for social content, with payouts settled directly through smart contracts.
Catch more Fintech Insights : The AI Shift in Fraud: Why Banks Need a New Playbook
[To share your insights with us, please write to psen@itechseries.com ]
