An API you can call today, and a protocol nobody can size

Thin day, and I am going to say so rather than dress two items up as five. The Cloudflare wallet went in Tuesday's edition, the FIS bank offering is from January and the Microsoft retail agents are from January too, all of which kept turning up in searches wearing 2026 URLs and none of which is news.

Two things did happen. One company shipped an API this morning that a developer can call. Another renewed the contract that quietly carries most of its business. And while checking a figure for the first of those I fell into a hole worth describing, because the most widely repeated number in this category is off by roughly a thousand times depending on which page you read.

OSL shipped something callable

OSL Group, listed in Hong Kong as HKEX 863, announced AgentPay this morning: stablecoin payment infrastructure aimed at AI agents making autonomous, intent-based payments [1].

The shape is familiar. A developer expresses an intent, meaning amount, asset and payee, and AgentPay handles routing, signing and settlement. It supports USDT, USDC and OSL's own USDGO among other stablecoins, and it speaks x402, AP2 and MPP among other protocols [1] [2].

Chief executive Kevin Cui framed it around distribution and liquidity, and William Yuan, who runs the AI Labs group, said stablecoins will become the base-layer asset for agentic economic activity [1]. Both are the sentences you write on launch day. Neither is why this one is worth ten minutes.

The reason is the protocol list. x402, AP2 and MPP are three competing answers to the same question, backed by different consortia with different economics, and OSL is supporting all three on day one. That is not diplomacy. It is a wager that the standards fight does not resolve this year, and it is an expensive wager, because every protocol you support is a settlement path you have to keep correct forever. Firms only take that bet when they think picking a winner early is the more expensive mistake. On current evidence they are right.

The caveat is the ordinary one and I will not bury it. This is a company announcing its own product. There is no independent volume figure, no named launch customer and no published pricing. What is checkable is that the API is open to integrate from today rather than in the coming months, which, after a week of wallets you can name but cannot fund, is a low bar worth clearing out loud.

One number in that release is not theirs and should not travel with their name on it: the McKinsey projection of three to five trillion dollars of global agentic commerce by 2030 [1]. It is a forecast in a press release, which is the least load-bearing object in finance.

Circle renewed the thing that actually matters

Circle confirmed on 6 August that it has renewed its long-term commercial agreement with Coinbase, the arrangement under which Coinbase distributes USDC [3]. Jeremy Allaire also ruled out paying quarterly distributions to USDC holders, saying Circle would keep reinvesting reserve earnings into the ecosystem rather than passing them to token holders [3].

The surrounding figures: USDC in circulation at $73.3bn, up 19 per cent year over year; on-chain transaction volume up 151 per cent; a reserve return rate of 3.5 per cent, which is down; and second-quarter revenue of $701.3m, up 7 per cent and short of what analysts wanted [3].

Yesterday's edition covered the agentic parts of the same quarter, so treat this as the other half of one disclosure rather than a fresh event. The half that went unremarked is the more revealing one. On-chain volume grew at more than twenty times the rate revenue did, 151 per cent against 7 per cent, because the reserve return is falling underneath the whole thing. Circle does not control that rate. It controls distribution, and distribution is what it just re-signed.

Context for why the renewal read as a relief: Coinbase, Visa, Mastercard and Stripe are all behind the Open USD consortium, and CoinDesk reported on 3 August that those backers still intend to support USDC alongside it [4]. Supporting both is cheap for them and existential for Circle, which is the whole asymmetry.

Read from the rails

Here is the hole. I wanted a current size for x402, since both of today's items lean on it, and I found two published figures that cannot describe the same object.

The protocol's own reporting, as carried in coverage of the Cloudflare launch, gives 75.41 million transactions and $24.24m of volume over a trailing thirty days, with 94,060 buyers and 22,000 sellers [5]. Elsewhere the protocol is reported as having processed $50bn across 200 million payments, a cumulative figure attributed to a Solana Foundation webinar recap [6].

Do the division, because nobody quoting these ever does. The first pair works out at about 32 cents a transaction. That is a micropayment rail behaving exactly as designed. The second pair works out at $250 a payment. That is not a micropayment rail, that is a wire transfer, and it is not what x402 is for.

Both numbers may be individually defensible. They are not describing the same thing, and one of them is almost certainly counting chain-level stablecoin movement rather than protocol-level agent payments. I could not resolve which from the sources available this morning, so I am reporting the discrepancy rather than picking the flattering one.

The practical version, if you are building on this: before an x402 figure goes into a board deck or a fundraising memo, establish three things. Whether it is a thirty-day window or a cumulative total. Whether it counts the protocol or the chain underneath it. And whether the average transaction size the figure implies is consistent with a protocol whose entire purpose is payments too small for a card. That last check takes one division and it catches almost everything.

I have watched numbers like this go unchallenged for months, because they are too big to question and too boring to divide. The people repeating them are not lying. They inherited them.

Thirty-two cents a transaction and two hundred and fifty dollars a payment cannot both be the same protocol. Somebody's denominator is wrong, and the number kept travelling anyway.