![[object Object]](/media/2026/09/hands-ai.webp?w=1920)
In the week of November 17, 2025, about 20 million stablecoin transfers moved over x402, a payment standard built for software that pays for things by itself. An AI agent reads a situation, compares options, and settles the bill without a person clicking anything, provided it stays inside limits its owner set in advance.
By 2026 the idea had left the demo stage. Coinbase, Google, Visa, Mastercard, Stripe, and PayPal have each shipped their own way for agents to pay, several within the same few months. A good share of that activity settles on public blockchains instead of bank rails, which makes it one of the more concrete tests of whether crypto has a job beyond trading.
What an agentic payment involves
An agentic payment is a purchase or transfer that an AI agent completes from start to finish for someone else: it picks what to pay for, checks the terms, and releases the money. The owner's control comes from boundaries drawn beforehand. Spending caps and approved recipients define the box, and the agent moves freely inside it.
x402 agent payments: the November 2025 peak is still unbeaten
Weekly stablecoin transfers over the x402 protocol, and which chains carry them
Transfers per week (week starting)
The Aug 17 week saw 8.7M transfers, more than double the week before, but only about 4 cents each.
Share of transfers by chain
Week of Aug 17, 2026
Week of Nov 17, 2025
Solana 6%, other 1%
Other in Aug 2026 is Polygon (1.1M transfers) and Algorand (109K).
View the data as a table
| Week starting | Transfers | Base | Solana | Other |
|---|---|---|---|---|
| Nov 17, 2025 | 20.1M | 93% | 6% | 1% |
| Aug 10, 2026 | 4.1M | n/a | n/a | n/a |
| Aug 17, 2026 | 8.7M | 48% | 38% | 14% |
Source: Token Terminal data, as reported by Cryptopolitan (August 2026). March 2026 weekly transfers were below 1 million (not charted, exact figure not published).
Chain shares are rounded. The Aug 17 week moved about $367,950 in total, roughly $0.04 per transfer.
Ordinary automation runs a rule somebody wrote by hand. An agent weighs context the rule-writer never anticipated, and that flexibility is exactly why the boundaries are worth getting right.
Why agents end up on crypto rails
Banks ask for identity documents and a human signature. Software has neither. An AI agent cannot open a checking account or plug into Fedwire, SWIFT, or ACH, which were built for people and institutions. What it can do is hold a private key, which is all a crypto wallet needs to operate.
That is why stablecoins, especially dollar-pegged ones like USDC, keep showing up as the settlement asset for agent activity. They clear in seconds anywhere in the world and can carry spending rules in code. BlackRock's head of digital assets, Robbie Mitchnick, has argued that agents will skip bank wires and ACH and settle in blockchain-based money instead.
The standards race: x402, AP2, and the card networks
No single protocol owns the space. Rival camps built different pieces, and real deployments usually stitch several together.
| Standard | Backed by | Rails it uses | Best suited to |
|---|---|---|---|
| x402 | Coinbase | Stablecoins over HTTP | Software paying software for data, APIs, and compute |
| AP2 (Agent Payments Protocol) | Card and crypto | A shared way for an agent to prove it was authorized to pay | |
| Visa and Mastercard agent programs | Visa, Mastercard | Existing card networks | Consumer purchases with identity checks and chargeback cover |
| Checkout inside ChatGPT (ACP) | OpenAI, PayPal | Cards, orchestrated | In-chat shopping across many merchants |
Card programs mostly cover consumer purchases, where chargebacks and merchant relationships already exist. Stablecoin rails cover the plumbing between software services, such as paying for an API call or a slice of compute, where speed and tiny amounts matter more than dispute handling. AP2 sits across both as a way to carry the owner's signed permission.
x402 and the return of "Payment Required"
x402, the standard from Coinbase, puts a long-idle piece of web plumbing to work. HTTP status code 402, "Payment Required," was reserved in the web's early days and then went unused for decades. Under x402, a server can answer a request by asking for a small stablecoin payment, and an agent can pay and receive the data in one exchange with no login or API key. Coinbase layered agent-specific wallets on top in February 2026, with spending controls built in, according to Spark.
Payments at this scale look nothing like consumer checkout. A single API query or a few seconds of compute can cost a fraction of a cent, which is why the protocol suits agents making thousands of them.
What the on-chain data shows
Transaction counts run into the millions while the money moving stays small.
- About 157,000 agents paid through x402 in the 30 days before Coinbase's June 2026 agent launch, according to x402scan data cited by CNBC.
- More than 165 million transactions worth roughly $50 million in total, across about 69,000 active agents, as of April 2026, per Spark. That averages out to about 30 cents a transaction.
- 8.7 million x402 transfers in the week starting August 17, 2026, up from 4.1 million the week before, but worth only about $368,000, or roughly 4 cents a payment. The figures come from Token Terminal, reported by Cryptopolitan.
- The all-time record still dates to November 2025: about 20.1 million transfers in one week, with just over $10 million in value, per the same report.
Activity is spreading, too. Solana's share of x402 transfers rose from 6% to 38% in nine months, according to that Token Terminal data, so developers are no longer defaulting to a single chain.
Read together, the figures show a market that spiked in late 2025, cooled, and is now rebuilding across more chains. Dollar values per payment stay tiny, so transaction counts are the better gauge of adoption for now.
Where the risk sits
Giving software spending power invites obvious worries: a drained wallet, a tricked agent, a runaway loop. The safeguards taking shape start with how the wallet is built. A scoped wallet holds only what the agent needs for one task, with a ceiling on each payment and a list of approved recipients, so a mistake stays small. Google's AP2 adds signed mandates, cryptographic records of what the agent was permitted to do, which give merchants and owners something concrete to check when a payment is disputed.
Public blockchains supply the other half. Every transfer is timestamped and permanent, so a post-incident review starts from facts rather than from logs someone could have edited. Many setups also send larger sums through a human approval step.
Monitoring tools built for human wallets are being pointed at agent activity as well, watching for a compromised agent or a payment heading to an address tied to known scams.
Telling agents and people apart
As agents multiply, merchants need to know a request comes from an authorized agent and not a bot running a scam. Several standards answer with cryptographic identity: the agent presents verifiable signed credentials instead of a shared password or API key.
The mirror problem sits on the human side. On a web filling with AI-generated activity, proof that a real, unique person approved something has value of its own, and blockchain-based identity projects that confirm someone is human and unique are pitching themselves as the answer. Identity, for agents and people alike, now matters about as much to this shift as the payment rails.
What could change the picture
Consolidation is the first thing to watch. If the standards interoperate, an agent authorized once could pay across many services, and much of today's friction disappears. Stablecoin settlement should keep growing underneath, though the August figures show it can stall as well as surge.
Regulation is the bigger unknown. Regulators are still working out how to treat a payment no human directly approved in the moment, and the answer will shape which wallets and chains agents are allowed to use. The consumer end may move ahead of the rules: MoonPay announced a virtual Mastercard for agents in May 2026 that spends stablecoins at any merchant accepting Mastercard, according to Spark's research.
Whether the dollar figures ever catch up with the transaction counts depends on who is willing to sign off on a payment nobody watched happen.
Where the risk sits
Giving software spending power invites obvious worries: a drained wallet, a tricked agent, a runaway loop. The safeguards taking shape start with how the wallet is built. A scoped wallet holds only what the agent needs for one task, with a ceiling on each payment and a list of approved recipients, so a mistake stays small. Google's AP2 adds signed mandates, cryptographic records of what the agent was permitted to do, which give merchants and owners something concrete to check when a payment is disputed.
Public blockchains supply the other half. Every transfer is timestamped and permanent, so a post-incident review starts from facts rather than from logs someone could have edited. Many setups also send larger sums through a human approval step.
Monitoring tools built for human wallets are being pointed at agent activity as well, watching for a compromised agent or a payment heading to an address tied to known scams.
Telling agents and people apart
As agents multiply, merchants need to know a request comes from an authorized agent and not a bot running a scam. Several standards answer with cryptographic identity: the agent presents verifiable signed credentials instead of a shared password or API key.
The mirror problem sits on the human side. On a web filling with AI-generated activity, proof that a real, unique person approved something has value of its own, and blockchain-based identity projects that confirm someone is human and unique are pitching themselves as the answer. Identity, for agents and people alike, now matters about as much to this shift as the payment rails.
What could change the picture
Consolidation is the first thing to watch. If the standards interoperate, an agent authorized once could pay across many services, and much of today's friction disappears. Stablecoin settlement should keep growing underneath, though the August figures show it can stall as well as surge.
Regulation is the bigger unknown. Regulators are still working out how to treat a payment no human directly approved in the moment, and the answer will shape which wallets and chains agents are allowed to use. The consumer end may move ahead of the rules: MoonPay announced a virtual Mastercard for agents in May 2026 that spends stablecoins at any merchant accepting Mastercard, according to Spark's research.
Whether the dollar figures ever catch up with the transaction counts depends on who is willing to sign off on a payment nobody watched happen.
FAQ
What are agentic payments?
What are agentic payments?
Payments made by software acting for its owner, with no person approving the final step. The owner sets the limits beforehand and the agent stays inside them.
Can AI agents spend crypto?
Yes. A wallet needs only a key, so an agent can hold and send stablecoins directly. Dollar-pegged tokens such as USDC are the usual settlement asset.
What is x402?
A payment standard from Coinbase that revives the dormant HTTP 402 status code. It lets one service charge another a small stablecoin payment over plain web requests, which suits agents buying data, APIs, or compute.
Are agentic payments safe?
Safety depends on the setup around the agent. Scoped wallets with payment ceilings and approved recipients, signed authorization records, and human approval for larger sums all limit the damage from a mistake, and public blockchains leave a permanent record of every transfer.
