jul 2026
i hit this properly for the first time a few weeks ago, and it was stupid.
a long running agent of mine needed something behind a paywall. the thing itself cost a couple of cents. moving a couple of cents through normal payment rails costs about thirty. so the cheapest correct action available to my agent was to not do the task.
thats not a pricing inconvenience, thats a wall. and once i started looking at it properly i realised the wall has been there the entire time, doing far more damage than i had ever noticed.
card rails cost roughly thirty cents plus a couple of percent. that fixed part is the important bit, because it means anything worth less than about thirty one cents costs more to sell than it is worth.
so nobody sold anything worth less than that. and the web, needing to make money anyway, went and invented every possible way around it. advertising. monthly subscriptions. free tiers that pay for themselves with your data. api keys with minimum commitments. bundles, because you cannot sell one article but you can sell nine hundred of them for five dollars a month.
we talk about those as business model choices. mostly they are workarounds for a number. an entire layer of the internets economics is downstream of a fixed fee on a card network.
agents dont create this problem, they make it impossible to ignore. a human reads four articles a month and grumbles about paywalls. an agent doing real work might touch fifty sources in a single run, and it cannot subscribe to fifty things, and the one article it needs from each is worth less than the fee to buy it. subscription pricing has no answer for that shape of demand. it was never designed for a buyer that wants one thing, once, from a thousand different sellers.
what changes with settlement on something like base is not that payments get a bit cheaper. its that the fixed part very nearly disappears. sending two hundred dollars of usdc on base costs about two hundredths of a cent, and it costs exactly the same to send two cents.
in practice you pay a bit more than raw gas, because most people settle through a facilitator rather than running their own. coinbase charges nothing for the first thousand settled payments a month and $0.001 each after that, which is a tenth of a cent, still flat, still indifferent to the amount.
so the floor goes from about 31 cents to somewhere between 0.02 and 0.1 cents depending on who settles for you. call it two and a half to three orders of magnitude of commerce that could not previously exist becoming arithmetically possible.
whether it does exist is a different question and i will get to it, because the honest answer right now is mostly no.
worth being precise here because i had it muddled myself.
x402 is a handshake. it takes http 402, a status code that has sat unused since the beginning, and gives it a meaning. your agent asks for something. the server says 402 and states its price and where to pay. the agent pays, asks again with proof, and gets the thing. four messages. thats the protocol.
what it is not is the rules. and the rules are the entire problem.
i had been sloppy about this, talking as though the rules ride on the money itself. they dont. stablecoins are fungible, there is no policy attached to individual units of usdc. policy attaches to accounts, allowances, session keys, spend permissions. thats a layer sitting next to x402, not inside it.
which sounds like a technicality and isnt, because it tells you where the real work is. the protocol part is nearly solved. the question of when should this agent be allowed to spend is not solved at all, and i assumed no standards body would ever solve it. i was half wrong about that, and ill come back to it.
the thing that kept me from giving my agents a wallet was never the cost, even though the cost is what i noticed first. it was that i had no way to say what i actually meant.
what i mean is something like: this agent, up to four dollars a day, only to these kinds of counterparties, only for this task, expiring on friday, revocable by me in one second, and every spend written down next to the request that caused it.
a card cannot express that. a card gives you a credit limit, merchant category codes, and a chargeback three weeks later. those are blunt instruments designed for a human who occasionally buys something, not for a process making four hundred decisions an hour.
a programmable account can express all of it, and enforce it at the moment of spend rather than reviewing it afterwards. thats the actual shift, and its worth stating in plain terms rather than reaching for anything grander: money that checks the rules before it moves, instead of money that moves and gets audited later.
i went looking for the competition and mostly found layers.
that was the genuine surprise. i expected a standards war and found a stack.
| what it is | which layer | what it costs |
|---|---|---|
| x402 — coinbase, now an x402 foundation project | settling one request | zero protocol fee. gas, plus a facilitator if you use one. coinbase’s is free for the first 1,000 onchain settlements a month, then $0.001 each, and one onchain settlement can be batched across thousands of payments |
| ap2 — google, sept 2025, now with the fido alliance | proving a human authorised it | no protocol fee. cards today, other rails on the roadmap |
| acp — openai and stripe, apache 2.0 | checkout | no protocol fee. you pay your processor. the 4% openai took on instant checkout ended march 2026 |
| ucp — google with shopify, etsy, target, walmart, wayfair | cart and catalogue across merchants | no protocol fee |
| stripe stablecoin | custody and settlement so you dont have to | reported at 1.5%. stripe doesnt publish the rate on the product page |
| card rails — visa, mastercard | all of it, expensively | a few percent plus a fixed fee around 30c, varying by country and card type |
| nevermined, skyfire, crossmint, kite | agent native billing, hybrid settlement | usage based, varies |
three things worth pulling out of that.
stripe didnt lose to x402, it adopted it. february 2026, usdc agent payments on base. coinbase then handed the protocol to a foundation under the linux foundation in july with visa, mastercard and google among the members, and aws and cloudflare put it at the edge. coinbase’s facilitator settles on base, polygon, arbitrum, world and solana, and there is a separate stellar implementation. whatever standards fight there was, it is over, and the incumbents won it by joining.
ap2 is the rules layer i said was missing. it works on signed mandates, a checkout mandate and a payment mandate, carried as verifiable digital credentials. in plain terms, a cryptographic permission slip a human issues before the agent spends, which a merchant can verify afterwards. thats precisely the thing i complained the protocol doesnt answer for you, and google standardised it in september 2025 while i was busy being annoyed about paywalls. the fido alliance is working on it now too, which is usually a sign something is becoming real rather than staying a blog post.
worth being accurate about where it actually is, though, because the write ups run well ahead of the spec. ap2 supports cards today. wallets, bank rails like upi and pix, and digital currencies are on the roadmap, not shipped. so the specific thing i want, one mandate that authorises a stablecoin micropayment, does not exist yet. i had read otherwise in three different summaries before going and reading the protocol docs, which is its own small lesson.
so the honest correction to my own complaint: the policy layer is not missing, i hadnt found it. what is still missing is the boring operational half. a mandate proves a human said yes once. it does not tell you how much budget is left, whether this counterparty is sane, or how to revoke everything at four in the morning. that part you still build.
the only genuine head to head is on cost, and it is a shape difference rather than a size one. stripe charges 1.5% to handle stablecoins whose underlying transfer costs about 0.00009%. thats not a rip off, youre buying compliance, custody and never having to think about a private key. but a percentage grows with the amount and the rail underneath it doesnt, so the two cross. above roughly a penny, the percentage costs more than settling yourself. below a penny, convenience is nearly free and you should just take it.
so the useful question was never which rail is cheapest. its where your transactions sit relative to that crossover, and how much you want to think about custody.
here is the part i want to sit with, because it is the strongest argument against my own enthusiasm.
the rails are real. the volume is not. x402 has processed 165 million transactions across 69,000 agents, which sounds enormous until you look at value: roughly twenty eight thousand dollars a day, and a good half of that is testing rather than commerce.
the standard is maturing much faster than the market that would run on it. and i dont think thats a temporary lag you can wave away, because there is a real chicken and egg underneath. sellers dont price per request, because nobody buys per request. buyers dont buy per request, because nothing is priced that way. a protocol existing does not make either side move.
theres a harder version of that objection too. the people who own the paywalled things may simply not want to sell one article for a cent. their business is the subscription, and per request pricing at a cent might be worse for them even if it is better for me. i can build the road and find nobody wanted to drive that way.
so i hold the position loosely: i think the floor moving matters enormously, and i cannot yet point at much that proves it.
there is no undo. i argued in the last memo that the decisions worth calling architecture are the ones that are expensive to reverse, and that the work is keeping as much as you can off that list. on chain settlement goes straight onto it. and i am proposing to hand the irreversible option to a system with a higher error rate than a human and no ability to explain itself afterwards. i dont have a clever escape. you dont get chargebacks back, and you cannot dodge that by routing through a big friendly processor either: stripes own stablecoin docs list dispute support as, simply, no. the chargeback does not exist on this rail whoever you go through. what you can do is bound the damage instead of undoing it: keep amounts small, use escrow where the counterparty is unknown, keep sessions short and revocable, and accept that the design goal shifts from fixing mistakes to making each one too small to care about. thats a worse guarantee than reversibility, honestly stated.
cheap payments break the fraud detection built for expensive ones. every anomaly system in existence assumes transactions are chunky enough to be noticed. ten thousand payments of a hundredth of a cent are invisible by construction, and they drain exactly the same budget as the one payment anybody would have flagged. the monitoring for this does not exist yet as far as i can tell, and it is not the same monitoring we have.
prompt injection stops being an information problem and becomes theft. today the worst a hostile page can do to my agent is make it say something wrong or leak some context. give that agent a wallet and the same attack spends money. untrusted input now touches value directly, which is a different category of risk from anything ive dealt with, and i say that having spent last week finding an unauthenticated endpoint and an ssrf in my own stack. i do not think i would find every hole in an agent that can pay.
because i think what counts as value is being renegotiated, and this is the layer where that argument gets settled in code rather than in essays.
for most of history the answer to what makes someone successful was some version of how much they had. if the optimistic version of the next decade lands, and thats the one i believe in, the cost of most goods falls hard, a lot of work goes away, and material scarcity stops being the thing that organises everyone. i dont think that ends competition. people dont stop competing when survival is solved, they just find something else to be scarce about. but the unit changes. attention, judgment, access, outcomes, whatever it turns out to be.
and machines transacting with each other is where that stops being a nice thought and becomes infrastructure. once agents can pay each other for work at a hundredth of a cent, value is moving between parties that were never in the economy before, in units nobody bothered to price, for reasons no human approved individually.
which brings it back round to the boring technical thing. whoever writes the policy layer is writing rules that non human actors obey at scale. not metaphorically. the budget, the scope, the counterparty list, the expiry. thats a small number of people deciding what a large number of autonomous things are permitted to want, and it is being designed right now, mostly by whoever gets there first, mostly without anyone treating it as the constitutional question it slightly is.
i find that more interesting than the transaction fees. and the transaction fees are what made it possible, which is usually how this goes. the boring number moves, and everything downstream of the boring number quietly rearranges.
so far there is a road and almost no traffic. ask me in a year.