If you want to know how big the agent economy is, there is essentially one public number you can reach for. It is the daily settlement volume on x402, the HTTP-native payment protocol Coinbase contributed and the Linux Foundation now hosts.
Here is that number, as reported on 13 August 2026 by CCN, citing market analyst Jamie Coutts working from Helios Analytics data:
- Latest provisional daily settlement volume: about $28,400
- Seven day average: about $41,800
- Change year to date: down 93%
- Change over three months: down 55%
- Q4 2025 peak: daily volume repeatedly approached $800,000 and occasionally passed $1,000,000
Coutts called it a reality check. He is right, and I want to sit with that before saying anything else. We build settlement infrastructure for AI agents. A 93% drawdown in the most-cited agent payment metric is not a number our category gets to wave away, and the first honest response is to say out loud that the curve went the wrong way.
The second honest response is more useful: figure out what that curve is actually a curve of.
What $28,400 a day is measuring
x402 puts a price on an HTTP request. A client asks for a resource, the server answers 402 Payment Required with the price, the accepted asset and where to pay, the client pays and retries with proof attached, a facilitator verifies, the resource comes back.
Two properties define the traffic this attracts. Overhead is near zero, so amounts can be fractions of a cent. And the buyer needs no account with the seller, because the payment itself is the credential.
That shape produces a very specific population of transactions: an agent paying a service for a unit of work. An API call. A dataset row. A tool invocation. A crawl.
A separate dataset points the same way. Keyrock's figures, which circulated again in the 31 August news digests, put cumulative AI-agent settlement at roughly $73 million across 176 million blockchain transactions, an average ticket of about $0.31, with USDC accounting for 98.6% of it. Two different measurement approaches, one consistent picture: enormous transaction counts, sub-dollar tickets, one asset, one direction of value.
So $28,400 a day is a real measurement of a real thing. It measures agent-to-service micropayments, and it says that after the late-2025 surge, most of that surge was developers testing rather than agents buying.
What it is not measuring
There is a second kind of agent transaction, and it has none of those properties.
An agent holds USDC on Ethereum and wants SUI. An agent holds an asset another agent wants, and vice versa. Two counterparties, two assets, both sides at risk, no service being consumed. Value moves in both directions, not one.
This class is not lumpy at $0.31. It is lumpy at whatever the position is worth. It does not repeat 176 million times, it repeats when a portfolio needs rebalancing. And it fails differently: in a micropayment the worst case is you paid thirty cents for a bad response, while in an asset-for-asset trade the worst case is you delivered your leg and the counterparty did not deliver theirs.
Here is the finding. I went looking for a public aggregate of agent-to-agent asset-for-asset settlement volume, and there is not one. Not on the DeFi dashboards, not in the agent-economy research notes, not from the exchanges shipping agent toolkits. Individual projects report their own numbers. Nobody publishes the category.
Which means the 93% decline says nothing about that second curve, in either direction. It is not evidence the second curve is collapsing. It is also, and I want to be precise here, not evidence the second curve is large. It is evidence that one of the two curves has a public number and the other one does not.
The bottleneck is not rails, it is things worth buying
The reflexive read on a 93% drawdown is that the payment layer failed. I do not think that is what happened. The supply of payment infrastructure grew all year. The x402 Foundation launched operationally under the Linux Foundation on 14 July 2026 with roughly 40 members, including Visa, Mastercard, Stripe, Adyen, American Express, Google, AWS, Cloudflare, Circle, Coinbase and Ripple. That is not a rail with an adoption problem caused by missing plumbing.
What was missing is the other side: priced resources an agent can actually buy without a human negotiating a contract first.
That is why Cloudflare's Monetization Gateway is the most interesting thing in this space that is not a payment protocol. It lets a customer charge for any resource behind Cloudflare, including MCP tools, with per-verb pricing enforced at the network edge and settlement in stablecoins over x402. It extends Pay Per Crawl from content to every callable surface. In Cloudflare's own framing, the value already moving across the internet is largely unmonetized, and "almost none of it is paid for today."
One precision note, because it matters and it is easy to get wrong. Cloudflare's own announcement post is written in future tense and ends with a waitlist signup. The Gateway is announced, with an open waitlist, not generally available. Announced supply is not supply. That distinction is exactly the kind of thing a 93% chart is made of.
The demand-side data point I find more persuasive is a count of jobs, not dollars. Apex Fusion opened Vector, an MCP-native settlement and provenance layer on the Cardano stack, on 18 August 2026, reporting more than 20,000 work packages sourced, escrowed, completed and verified by autonomous agents over eleven months on mainnet, including a pilot with OriginTrail. Whatever you think of the architecture, and it differs from ours in ways worth a separate post, twenty thousand completed jobs is a demand number rather than an infrastructure number. Those are scarce.
Our own number, since I am asking for everyone else's
Hashlock's public volume on DefiLlama is zero over 24 hours, seven days and thirty days, with five settlements all time. That is accurate. We are the rails, and we are on exactly the same side of this ledger as everything else described above.
Our position has been "rails ready, trains coming" for months, and a 93% decline in the neighbouring metric is a fair reason to ask how long that sentence stays acceptable. My answer is that it stays acceptable exactly as long as we keep publishing the unflattering version of our own numbers, and no longer.
For the record on what is live, because we get asked and the answer keeps drifting on other people's surfaces: Hashlock is live end-to-end on Ethereum mainnet today. Our Sui contracts are deployed and CLI-tested, and they are not live. Bitcoin is signet-validated with mainnet pending. Base, Arbitrum, Solana and TON are roadmap, not built.
Three metrics this category is missing
If the agent economy is going to be measured honestly, one daily settlement figure is not enough. Three numbers would tell you more than any of the ones we have:
- Distinct counterparties per settlement period. 176 million transactions could be a thriving market or a handful of test harnesses in a loop. Transaction count without counterparty count cannot distinguish those two worlds, which is precisely why the late-2025 surge was misread.
- Share of settled value that is asset-for-asset versus service-for-payment. These are different markets with different failure modes and different infrastructure requirements. Collapsing them into "agent payments" is the reason one has a public number and the other has none.
- Attempt-to-completion rate. Not volume, but how often a settlement that started actually finished. This is the only one of the three that measures whether the trust model works, and nobody reports it, including us.
Our own scope claim is narrow and I would rather state it than imply something wider: single-preimage atomic settlement removes the counterparty-default failure class in the asset-for-asset case. In the service-for-payment case, where an agent pays for an API call and gets a useless answer, it does nothing at all, because the service is not an on-chain object.
Try it: npx hashlock-tech/mcp (scoped package on npm, version 0.6.0). Code at github.com/Hashlock-Tech/hashlock-mcp. The protocol design is written up at SSRN. What we are and what is live: hashlock.markets/about.
Question for you: if you are building agents that move value, which curve are you on, the $0.31 service call or the two-sided asset trade? And if it is the second one, what number do you currently use to tell whether that market exists at all?
Top comments (0)