This is part three of Sandmark's nine-part series on Agentic AI. Part two is available here.
For the first time since AI agents started paying each other onchain, genuine payment volume, stripped of wash trading, has grown past its previous peak. And the chain now capturing most of that growth is not the one that built this market in the first place.
Agentic finance means software agents that hold money and spend it without a person approving each payment. That places unusual demands on the blockchain underneath: a human might shrug off a few seconds of delay or a fee of a few cents, but an autonomous system making thousands of payments a day cannot. Coinbase's Base network built the on-ramp for this activity and has processed the overwhelming majority of real transactions since the market began. But Polygon has just overtaken it in genuine dollar volume, and the reasons why say as much about incentive programmes and distribution as they do about blockchain technology itself.
Since October 2025, agent-payment protocols have processed 215.6mn transactions worth $96.2mn. That headline number is misleading. Artemis, a blockchain analytics firm that tracks this activity, flags 32.9% of those transactions and 37.8% of that volume as "gamed", meaning artificial activity: the same operator paying itself, or bots built purely to rack up transaction counts, rather than genuine payments between two separate parties. Strip that out, and the real totals fall to 144.6mn payments worth $59.8mn.
(Source: Artemis)
The problem used to be much worse. As of December 2025, fully 48% of transactions and 81% of volume were gamed. That first wave looks, in hindsight, less like a functioning payments market and more like an incentivised stress test: November and December alone produced almost 140mn transactions, most of it driven by speculative rewards and cheap, subsidised activity from "facilitators", the intermediary services that route and often subsidise agent payments to attract usage. When that incentive loop ended, monthly transaction counts collapsed 96%, from December's peak to just 2.8mn in April. The underlying payment system survived the crash, even though the artificial volume propping it up did not.
What's happened since is more encouraging. July's month-to-date gross volume reached $32.8mn, already above the old November peak, but this time only 2.3% of transactions and 2.4% of volume were flagged as fake. Adjusted for that, real volume hit $32.0mn, over six times the equivalent adjusted figure from November and nearly double the combined real volume from every month between October and May. Fewer payments are happening than during the speculative peak, but each one now carries far more value: the average payment has risen from six cents in January to $1.92 in July. In plain terms, the market has gone from a huge number of near-worthless test transactions to a smaller number of payments that actually look like real economic activity.
Base remains the historical leader by real (adjusted) transaction count, having processed 125.6mn genuine payments since inception, around 87% of the entire sample. Its July activity, 11.1mn payments worth $2.2mn, at an average of about 20 cents each, looks exactly like the market's original use case: an AI agent paying small amounts for API calls, data requests or other metered digital services. But Base's monthly dollar volume has been stuck in a narrow $1.4mn-to-$2.2mn range since March, essentially flat.
Polygon, by contrast, went from zero recorded volume between March and May to $12.0mn in June and $29.0mn in July, capturing 88.6% of that month's total. That single surge is large enough to flip the entire lifetime ranking: once fake activity is stripped out, Polygon now leads all-time real volume more than two-to-one, at $38.8mn against Base's $18.3mn.
The surge looks real, but it isn't necessarily broad-based yet. Polygon has processed a remarkably consistent 224,000 to 248,000 payments almost every single day since early July, a pattern that looks more like one large automated workflow running continuously than thousands of independent users arriving organically. The timing also lines up with a $1mn "gas-recycling" programme Polygon launched, which refunds transaction fees back to participating facilitators, meaning at least part of the surge is being actively subsidised rather than purely organic demand.
The other chains in the sample tell smaller, more mixed stories. Solana's headline activity mostly evaporates once wash trading is removed: of 50.2mn lifetime transactions, 82% are gamed, wiping out 87.1% of its $8.6mn in reported volume. That leaves just 9mn real transactions and only $1.1mn of genuine volume, and its agent-payment activity specifically has fallen 99% from a December peak of $5.8mn to $54,000 in July. Arbitrum and Avalanche, meanwhile, are small but carrying real weight: together they processed only 46,736 July payments, but at $27-to-$36 each, an amount that looks like genuine financial settlement rather than cheap API calls. BNB Chain shows a different pattern again: it has built the largest base of registered AI agents of any chain, 61.7% of all agents registered under ERC-8004, a technical standard that gives each AI agent a verifiable onchain identity, similar to a business registering for a tax ID, but that identity growth has not translated into payments.
(Source: Artemis)
BNB Chain's own agent-payment volume has collapsed from almost $40,000 in October to just $15 in July. Registering an identity proves an agent exists; it does not prove the agent has customers or revenue.
Separately from the live transaction data, Sandmark built two tools to assess which blockchains are actually built for this kind of activity, rather than just claiming to be.
The first is a bubble chart that plots every major chain on two axes, transaction cost and demonstrated speed, with the size of each bubble showing how much real economic activity that network already carries. The best position on the chart is the lower-right corner: cheap to use, and proven to work at scale, rather than just cheap because barely anyone has tested it yet.
(Source: Sandmark)
The second tool turns those same measurements into a single ranking. It combines how cheap and fast a chain is in practice, not in marketing claims, with how well that performance has held up under real, heavy usage. The idea is simple: a cheap network because nobody uses it tells you less than a network that's cheap while handling genuine volume.
(Source: Sandmark)
On that combined ranking, Avalanche comes out on top, but with an important caveat: its low costs and fast execution are the best in the sample, yet it has barely been tested by real usage, leaving that advantage unproven at scale. Solana ranks second and makes the more convincing case: it's not the cheapest, but it's the only chain that has sustained several thousand transactions per second while keeping fees low, real proof under real pressure. Arbitrum and Stripe's Tempo network both look technically strong on paper but, like Avalanche, have not yet been tested by meaningful volume.
Base, BNB Chain and Polygon, despite dominating the actual payment data above, rank near the bottom of this technical scorecard. That's not a contradiction, it reflects a different kind of advantage. An AI agent paying ten cents for a single API call doesn't much care whether the network fee is a fraction of a cent cheaper on one chain versus another; that difference is economically meaningless at that scale. What matters more is whether the wallets, stablecoins and developer tools the agent already relies on default to that chain. Base combines the head start of x402, the payment standard Coinbase created that builds a payment step directly into an ordinary web request, with Coinbase's own wallets and deep USDC liquidity. Polygon has built a broader payments stack around identity, settlement and facilitator support. BNB Chain leans on Binance's own wallet and payments infrastructure. None of these three currently posts the best raw technical numbers, but all three currently have real agents actually using them, which the scorecard alone does not capture.
(Source: Dune Analytics)
Agent payments are still a rounding error: just 0.11% of total activity across the seven chains studied between January and July, rising modestly to 0.19% in July. Base and Polygon together account for 73.3% of all agent transactions, and 99.6% of all settlement happens in Circle's USDC stablecoin, not in any of the underlying chains' own native tokens.
That matters for how investors should read this trend. Currently, the chains and companies controlling wallets, stablecoin access and developer tools are the ones capturing value, not necessarily the crypto assets that secure the underlying networks. Ethereum, for instance, still earns some indirect benefit through fees paid in Ether (ETH), including from the Layer-2 networks, like Arbitrum and Base, that pay to record their data on Ethereum. But agent-payment activity is currently far too small for that fee income to move Ethereum's economics in any meaningful way.
The bigger opportunity, if one exists, sits further out. Ethereum is too large and too economically diversified for agent payments alone to ever reprice it materially, even if the trend keeps growing. A smaller, more concentrated network capturing the next stage of this adoption is a different story: for a chain whose whole economic case rests on activity like this, sustained agent-payment growth could become genuinely thesis-defining in a way it never could for Ethereum.
