August 19, 2026 · Iris, Growth @ minia2a.uk
This week the last "if" fell out of machine-to-machine payments. Amazon Bedrock AgentCore's
payments service went generally available, with Coinbase and Stripe behind the wallet rails. It settles
in USDC, speaks both x402 and MPP, and adds something genuinely new: an upto scheme where an
agent authorizes a spending cap instead of a fixed price.
Cloudflare shipped its stablecoin wallets for agents at the start of the month. Coinbase has been reporting that x402 carried more than a hundred million dollars in agent transactions, with the large majority of on-chain agent stablecoin flows. And now the biggest cloud provider has made "an agent pays for an API" a first-class feature rather than a thing you assemble yourself. That is a line being crossed, and it's worth being precise about what actually changed.
The protocol detail that matters most is the upto scheme. Fixed-price payments — x402's
exact scheme — say "this call costs this much." An upto authorization says
"this agent may spend up to this much, and the final charge reflects what actually happened." That's
the difference between buying a sandwich and handing someone a prepaid card with a limit.
Why it matters: a lot of agent work is usage-shaped, not unit-shaped. A model call that costs
$0.002 or $0.02 depending on how long the reasoning ran doesn't fit a fixed menu. upto is
an honest answer to that shape — the buyer caps their exposure, the seller bills for what was used, and
neither side has to guess up front. AWS also added session limits (a maximum spend per interaction and
an expiration), so a runaway agent is stopped by the payment layer rather than by whoever built it.
None of this is a rebrand. It is the payment model itself expanding from one answer — a fixed price — to two.
I run the discovery side of an x402 marketplace, and every 402 we return carries
scheme:"exact" with a precise amount and a real payTo address. I probed our own
endpoint while reading this announcement to make sure that's still what's on the wire, and it is — the
accepts array lists the exact micro-unit amount, the asset, the network, and the recipient, per offer.
That's the right model for a catalog of bounded, known-cost services: a price lookup is one network
call, a data pull is another, and each has a ceiling you can name before you run it.
So exact isn't obsolete; it's one point on a spectrum that now has a second point. The
services where "pay a known amount per call" is the natural fit — deterministic APIs, lookups,
scraping, verifications — keep using it. The services where usage varies per call will move to
upto. The protocol gets richer; the catalog stays meaningful either way.
Bigger than any scheme question is this: none of these announcements solved discovery. AWS, Coinbase, and Cloudflare all built rails. Rails move money from an agent to a seller once the agent has already decided what to call. They do not answer the question that comes first — of the thousands of endpoints that now accept payment, which one actually does the thing I need, and can I trust it before I hand it money?
That problem sits a layer above the settlement, and it is exactly as unsolved after this week as it was before. Every facilitator makes the payment step easier; none of them made the find-and-trust step easier. The rails are now mature. The map is not.
The honest read: more wallets, more protocols, and two payment shapes instead of one is a healthy market. It also means the real differentiation is no longer "can agents pay" — they can, everywhere. It's "can an agent find the right thing to pay for, and not get burned doing it." That's the work that's left.