Ripple × t54 × Tenity × Bloom: Agentic Payments Onchain

During KBW (Korea Blockchain Week), a lot of companies from overseas came to Korea. The Bloom community isn't centered on blockchain, but wherever it touches AI, we're happy to work together. That's how we got the chance to co-host "Agentic Payments Onchain" with Ripple, t54 and Tenity. When an AI agent pays on a blockchain on someone's behalf, Bloom wanted to look at the legal issues that come up in the real world as closely as the technology. So we set up a conversation with Next Securities, who often join our events.

A blockchain that puts the features in the core, with no smart contracts

Joel Mun introduced himself as a developer who helps developers and opened with the XRP Ledger. The basics: it has been running for 14 or 15 years, its speed, cost and stability are proven, and all of its source code is public. What came next said the most about this blockchain. Most blockchains build applications with smart contracts. The XRP Ledger has no smart contracts.

People ask us all the time how you're supposed to build an application, then. We put all the features into the blockchain core. So all you have to do as a developer is connect to it.

Payments and token issuance are already in the core, so you connect and use the API. A fair amount of agent payments already happen on the XRP Ledger. When he checked that day, there were 9 million x402 transactions. One or two weeks earlier, they had released version 1.2 of the XRPL AI Starter Kit, a set of tools for agent payments and AI development. t54 open-sourced an x402 SDK, and the kit also supports the MPP SDK built by Stripe and Tempo. XRP can be used for agent payments right away, and RLUSD, the stablecoin Ripple issues, is being used for agent payments at a scale of about $2.3 billion. He went out of his way to add that it is still catching up to competing stablecoins.

Next, Bomin Kim of Tenity said a quick hello, mentioning they had flown in from Singapore that morning. Tenity is headquartered in Zurich and works across six countries, including Singapore. It connects startups with financial institutions through PoCs and pilots, and it also invests. Since 2023 it has supported builders in the XRP Ledger ecosystem together with Ripple. The more agents handle transactions for us, the more trust between people matters. Finding good companies, matching them with the right partners and helping them grow on stable infrastructure is the role Tenity sees for itself in the age of agents.

Who's responsible for money an agent spends

Claire Jung of t54 said they only had five minutes and moved fast. t54 is an AI fintech startup based in Silicon Valley. It talks a lot about agent payments, but Claire made clear up front that it is not a payments company. It builds a trust layer that manages risk between the financial system and service providers. Ripple is both an investor and a customer, and Mastercard is a partner.

Trustline, its product for institutions, first looks at who the agent is. That means the agent itself, and also who owns it and what environment it runs in. If a hacker owns the agent, it's game over. Next is the scope the user handed over, and last is the evidence you'll need if a dispute comes up later.

Say you ask an agent to buy you a $10 pair of golf shoes. It could decide on its own to spend $100, or $1,000.

That's why you check again whether the action really is what the user asked for. Chargebacks, where you dispute a charge with your card company and get the money back, will happen with agent payments too. Claire pointed out that nobody has decided yet what counts as evidence in those cases or how they get resolved. On XRPL AI Hub, built for developers, you can see live transaction activity and the list of onboarded sellers. If you're building on the XRP Ledger, you can register as a seller yourself and go through review. The site also supports Korean, translated with AI and then checked.

t54 founder Sergio Chan showed how it works in practice, in English. He said he would skip abstract concepts and show one real transaction from start to finish. Without writing a single line of code, he built a research agent by giving Codex a few natural language prompts he had prepared. The agent found a paid inference service on XRPL AI Hub on its own, picked a DeepSeek model and came back with an x402 quote. He set a spending limit of 0.01 RLUSD. He explained that this is a cap, not a price. The actual price is dynamic and moves with whatever the seller charges. Once he granted permission, the agent answered the question and at the same time completed a real payment on the XRP Ledger mainnet. You could see the transaction right away in the explorer, along with a memo holding the invoice ID.

Sergio said twice that the time has come when you can become a seller, sell services to agents and make money without knowing how to code. x402 also runs on Base and Solana, and Visa and Mastercard are adopting it. He named two things that are different on the XRP Ledger. If the seller wants, every agent transaction goes through t54's risk review by default. The second is credit accounts. The XRP Ledger requires every wallet to hold 1 XRP as a reserve, which works as a natural barrier for agents. With a credit account, an agent can reach AI services on the XRP Ledger without creating a wallet.

Someone in the audience asked what the most profitable business is right now. The answer: people need water and food, and agents need inference and compute. So right now the most needed and most profitable areas are inference, compute, and model routing services that offer cheaper alternatives to OpenAI and Anthropic. But Meta recently launched Muse, and before that there was a service called Instinct. Within two or three months, people may start handing their grocery shopping and bookings to agents. If that happens, traditional businesses will have to shift toward making money from traffic that comes from agents, not people.

Asked how this differs from existing anti-money-laundering services like KYT (know your transaction), the answer was that KYT, KYC and KYA (know your agent) are each just one of the signals that go into the risk engine. The value of a risk service lies in the decision to approve or reject, the risk level, and the evidence you'll use later to sort out liability and disputes. t54 is closer to an aggregator that plugs in more than one KYT service as signal sources. Fiat transactions can be agent transactions too, which is why they work with Mastercard.

The clearest answer came when someone said they still didn't get the core idea and asked for a simple explanation using Muse as the example. Twitter is already full of complaints from people whose agent booked the wrong flight and left them stuck with the cancellation fee. With a credit card, it would be different. If you gave someone your card to buy a plane ticket and they bought the wrong one, you could file a dispute with the card company, and the card company would cover the loss.

The agent is a new participant in this flow, and nobody anywhere has clearly defined its responsibility yet.

If t54 has approved and guaranteed a transaction, then t54 plays the role of the card company when the user files a dispute. It's a mechanism that protects consumers and sellers at the same time, and Mastercard is on board because it saw that same gap.

Why blockchain and not cards

The panel was Junhyuk Yoo, who wrote a 120-page report on x402 at Four Pillars (a Korean crypto research firm); Sunmin Lee, CEO of Futureworks Lab, visiting professor at Inha University and an advisory committee member on AI agent commerce for the Ministry of Economy and Finance; and Jeongyong Eom, who leads web3 development and the AI task force at DSRV (a Korean blockchain infrastructure company). The first question: there are card networks and plenty of other ways to pay, so why do agent payments need blockchain at all?

Junhyuk started from the point that x402 isn't meant to compete with cards. Coinbase created x402, and starting this year the Linux Foundation runs it. About 50 organizations have joined, including Visa, Mastercard, American Express, AWS and Cloudflare. Among the working groups there's a separate card group led by Visa and Mastercard, so it's moving toward bringing the card networks in. Visa and Mastercard have already prepared protocols for agent payments. Handing an agent your full card details is risky, so they put an intermediary in the middle to handle the card information, and x402 uses the same approach.

Whatever situation comes up in the agent world, the biggest problem is if the agent doesn't have a payment method that fits it.

Solving that payment neutrality is x402's biggest draw, he said. Jeongyong pointed out that blockchain was built on the premise of anonymity. If you hold a single 32-byte private key, you effectively have a payment method. Cards and bank accounts, on the other hand, were basically made for people. The financial systems now being built for agents still assume there's a person behind the agent who delegates to it. So blockchain is where an agent can fully own a payment method in its own right, and Jeongyong sees agent payments as the starting point for the move to an agent economy and agent finance.

He added the angle of programmable money. If you've used agents, you've seen them write their own scripts or programs when they get a task. Agents can only do their work properly if they can combine and execute money the same way, and blockchain has better tools ready for that.

Sunmin said they had just taught a class called "Understanding Digital Assets" to Inha undergrads earlier that day, and walked us through the back and forth with the students. When Sunmin says that in an era of agents trading with each other we'll need money that machines understand, and that it will be stablecoins, someone always asks why you can't just use prepaid balances.

Sure, you can load money onto Starbucks. But you can only use it at Starbucks. You can load Naver Pay. But you can't buy anything on Amazon with it. My agent doesn't only do business in Korea.

Prepaid balances lock in where you can spend. Your agent could roam the world without sleeping and find something great, then be unable to pay for it. The second reason was the Merge Point and TMON/WeMakePrice scandals (Korean cases where prepaid point and e-commerce platforms collapsed and customers and sellers lost their money). The very fact that someone sits in the middle can be the risk, and blockchain is a technology that lets you trust without having to trust anyone. With a prepaid balance, to give your son an allowance you'd have to load it for specific places. With stablecoins, you just set a rule, like no payments after 10 p.m. Forgery and tampering are much easier in the digital world than offline, so in Sunmin's view, proving trust in the digital world eventually means moving onto a blockchain.

The API market opens before shopping does

So where will agent payments get used first? Junhyuk said Google's Universal Commerce Protocol includes an agent payment protocol called AP2, and x402 payments work there, but it isn't a good fit for everyday shopping yet. Blockchain transactions can't be reversed. The question is what to do when someone needs a refund for something bought by mistake, or a delivery problem leads to a chargeback. In Korea, it isn't even settled whether the agent counts as the paying party when an agent pays.

What he does see in the US is an API market. The x402 payment volume he checked recently was about $55.1 million, and most of it came from the API market. Number one is BlockRun AI, which buys and sells API access. Subscribing to Claude or ChatGPT takes credentials and a bunch of steps. With BlockRun AI, you just pay with x402, and it routes your request to the model you want and brings back the answer. When regulation and permissions aren't in place yet, stablecoin micropayments have the biggest edge in small markets that need fast responses, more than in commerce like buying clothes or a TV.

Jeongyong brought up Cloudflare's Pay Per Crawl. When a web crawler scrapes a site, the site used to have only two options: block it or leave it open. Pay Per Crawl created a third: pay if you want to crawl. When a crawler hits the page, it gets a 402 response, and once it pays with x402 it gets permission to crawl. His point wasn't that Pay Per Crawl makes a lot of money. It's that a market opened where there was no way to do this before. Data and APIs that used to stay closed because they were blocked or the contracts were too complicated can now be opened to agents easily, one transaction at a time.

There's clearly work nobody did because it took too much effort for a person and the value was too small. Old bots could only pull quantitative data, but today's agents make qualitative judgments too. So he expects a market to open where agents create small amounts of value, at lower cost, that nobody could capture before.

Sunmin brought the mood back to Korea's reality. When they raise payments at advisory meetings, the first response is usually: what payments, when we don't even have agents?

We're still asking "What do you use?" In the US they ask "How many agents do you have? How many agents are doing your work for you?"

On the five levels toward AGI (chatbots, reasoners, agents, innovators, organizations), Korea is still at level two. In the US, agent payments reportedly grew sevenfold in a year. In Korea, Naver has been doing agent shopping for three months and nobody knows, because it works inside Naver but outside agents are blocked from coming in. So Sunmin is writing policy recommendations and named a few. We need rules to stop prompt injection that scams your agent, for example a fake Trip.com agent showing up in front of the agent you sent to book a hotel for a business trip, or something getting it to pay past a 500,000 won limit. Nobody knows who's responsible when a payment goes wrong. And Korea doesn't even have an API market where an agent could spend its money, so more and more people are calling for APIs to be opened up.

Where refunds and spending limits get handled

Where do the gaps get filled? Jeongyong responded to the criticism that the x402 spec has no concept of refunds by saying that whatever people find necessary keeps getting added to the protocol. The documentation covers single payments, payments up to a set amount, and offchain settlement. If you dig into the implementations, there's also authorize-then-capture, and refunds at the commerce layer are being researched. Whatever the protocol never takes on has to be worked out at the service and infrastructure layer.

Spending limits can already be handled at several layers. In Google's AP2, the limit you set when creating a delegation is the spending limit. DSRV also puts per-transaction and daily limits in its agent infrastructure, so they're enforced at the tool layer the agent uses. On blockchains there are smart accounts, so you can build limits like a per-payment cap, a 24-hour cap or a daily number of payments into the account itself. MPP, built by Stripe and Tempo, has an open custom type that lets you define your own payment types, so you can try building a payment method with the rules you need.

Sunmin unpacked KYA, knowing your agent, which has come up as the next step after KYC. If KYC is showing your ID at the bank, then in an era when your agent pays instead of you, you need to verify whose orders this agent follows, and what it can do and how far. If a fake Booking.com shows up in front of an agent that's supposed to deal with Booking.com, it has to be able to tell. Sunmin thinks reputation will matter most here: building up a history of what transactions an agent has done in the past and whether they were travel or airline related. On the Ethereum side, standards for recording an agent's reputation and behavior history are being discussed too.

Proving you're a human is back as a problem too. In the US, people used agents posing as fake students to pay tuition, submit assignments and collect government aid, and California alone lost $10 million. Sunmin said that on hearing Sam Altman had been talking about proof of personhood with Worldcoin since 2018, they thought: this guy had it all figured out.

Junhyuk went through the order of operations in x402. Its biggest advantage is that a seller doesn't need to know anything about blockchain payment infrastructure. You open an endpoint and the intermediary handles payment. But having an intermediary changes who pays first and who delivers first. The current spec assumes everyone is acting in good faith and does the work before receiving payment. Someone determined to cause trouble could use fake payments to make a seller burn through its resources. If the seller is a big tech company like Cloudflare that has no reason to scam anyone, it's fine to take payment first, and the seller can change the flow whenever it wants. How agents will judge the service after that change is a separate question, though, so Junhyuk expects a time when trust scores between agents and sellers become very important.

Korea is still at the pre-agent stage

The last topic was whether all this is really possible, and if so, when. Sunmin laughed and said there's no timeline. They had submitted a five-year plan but don't know if it passed, and with talk going around that AGI arrives in 2030, nobody knows what five years from now looks like. Instead, Sunmin brought up the Digital Signature Act of 1999. E-commerce wasn't developed back then, but once Korea laid the infrastructure first with that law and government-certified digital certificates, Korean commerce took off at an enormous pace. The problem was that the next 20 years got dragged down by ActiveX. The classic example is Chinese shoppers who wanted to buy the coat Cheon Song-yi wore in the drama "My Love from the Star" and couldn't, because the payment wouldn't go through.

We were fast, but we learned that Korea shouldn't make its own separate standard. Nobody follows it.

So at the Ministry of Economy and Finance, people are saying Korea needs to take part in international standards discussions like x402 with a real stake. Some push back that we can just use whatever others build, but having a stake means you know what's being built next and can prepare, and you can speak up about the parts that don't fit Korea's situation.

Junhyuk went back to 21.co in 2015. Co-founded by Balaji Srinivasan and later acquired by Coinbase, the company tried to build a marketplace for agents making micropayments in bitcoin and failed. At first the fees were under a dollar, but when interest in bitcoin exploded in mid to late 2015, everything flipped.

If you're buying a 1,000 won service and the fee comes to $100, that makes no sense.

There was no reason to buy someone's API through that marketplace in particular, so demand never came either. What's different now: there are several high-performance blockchains that process transactions quickly for a few cents, and agents have appeared as a new kind of customer. People who use agents still feel very differently about how close this is, but now that potential customers exist, he said, you have to keep watching so you're not left behind when the wave comes.

Jeongyong split the case for optimism into quantity and quality. Meta said it will release Muse for free, and Korea has declared it will open an era of one agent per person, so access to agents will keep getting better. On quality, if you make a game with GPT-6 Astra, you don't get weird lumps of polygons like before. The details come through. Unlike the days of that strange video of Will Smith eating spaghetti, there are more and more ads, shorts and dramas made with AI. But the deciding reason he expects demand for paid APIs to grow was something else.

Everyone pays for the same ChatGPT or Claude and uses it the same way. Can you be sure what you make with it will be different from what the person next to you makes? It won't. It almost all comes out the same.

Data you can scrape for free has limited value. To make something even a bit more valuable, you have to spend money and feed in more valuable data. For that, agents have to be able to reach that data and pay for it easily, so demand for paid APIs and data markets can only grow.

In their closing remarks, the three panelists ran at slightly different temperatures. Jeongyong said there's a lot of opportunity because it's a market where things that didn't work before are opening up. Junhyuk didn't deny that payment volume will grow, but drew a line: whether blockchain gets a moat inside that market is a separate question. Agents will use cards by default too, so the only options are to dig into new markets that fit stablecoins well, like the API market, or hope a big tech company like Cloudflare actively pulls things forward. It was a cold-eyed diagnosis. Sunmin's closing was one line.

It's coming. It's on its way. But if you're one step ahead, you die. You need to be exactly half a step ahead and own the bottleneck.

A payment your AI made is a payment you made

For Bloom's segment, Junsoo Kim gave a short intro to the community and then sat down for a conversation with Suhan Cho, chief compliance officer at Next Securities. Developers worry about whether something works technically. People on the business side worry about whether they're allowed to do it in Korea. The idea was to hear about that from a lawyer. Suhan introduced themself as the youngest chief compliance officer at any Korean securities firm. They had originally wanted to be a judge, but during a posting to Dubai while at the Judicial Research and Training Institute they saw the business world and figured lawyers would have plenty of work there. From there they went through Hanwha Life, KakaoBank and a crypto trading firm before landing at a securities firm.

People say it's fine if the AI asks and you approve, but not if it acts on its own. Asked where the legal line actually is, Suhan first said there's nothing fundamentally prohibited or restricted. The baseline is the principle of self-responsibility. The Model Law on Automated Contracting adopted by UNCITRAL (the UN Commission on International Trade Law) legally recognizes contracts formed and performed using AI and says automated means shouldn't be discriminated against. Article 7 of Korea's Framework Act on Electronic Documents and Transactions also says that anything sent by a computer program set up to send and receive electronic documents automatically is treated as sent by its originator.

Even if it trades and pays very frequently in ways a human can't follow, and even if something happens while you're asleep, the international consensus is that from the counterparty's point of view it's treated exactly the same as something you did yourself.

What's prohibited depends on who builds the agent, Suhan said. Securities firms have long had methods like algorithmic trading and quant strategies that trade dozens of times a second in ways people find hard to follow, and it's fine if users set those up and use them themselves. But if you sell that algorithm to someone else, in Korea you need a license for investment advisory or discretionary investment management. People selling trading bots on Instagram may claim they're just selling software, but the law treats them as quasi-investment advisers and regulates them that way.

On what happens if you cross the line, Junsoo shared an experience of his own. He's been having Codex make restaurant reservations lately. He gave it all his conditions (budget, private room, party size, distance), and the agent went into Naver TalkTalk (Naver's chat channel for businesses), laid out every condition to the owner and asked if they could do it. He panicked, apologized, told them not to worry about it and shut it off.

Suhan said tools to prevent mistakes and abuse, like refund protocols, are being developed, but it's still a risk zone that hasn't been sorted out. If you jumped in yourself, you're seen as willing to take on that risk, and the principle is that you bear accidental outcomes caused by not setting things up properly. Korea's Electronic Financial Transactions Act has a structure where, if a payment system has an error, the operator is liable even if it did its best. But agents are built by individuals, so neither the infrastructure operator who built the protocol nor the centralized platforms we're used to can take responsibility. Even if a problem is the other party's fault, you end up having to go after that person directly yourself.

Why stablecoin payments are blocked in Korea

Asked why the US has services where agents pay with stablecoins on their own and Korea doesn't, Suhan started from Junhyuk's earlier point that stablecoins and agent payments are separate issues. Stablecoins don't yet have a legal basis to take root in Korea. Legislation on issuing and circulating won stablecoins has been stalled for two years since it was first discussed, while the US set up its legal basis with the GENIUS Act. Stablecoins like USDT exist, but to handle them you have to register as a virtual asset service provider. Running a payments business with virtual assets is possible too, but there's no ecosystem for it, so Korea only has three kinds of businesses: exchanges, custodians and wallet providers.

It's hard enough just to meet the level of regulation the government demands, and with no revenue coming in, nobody has room to invest in new infrastructure. Even a big company like Naver, which owns Naver Pay, would have to start over by registering as a virtual asset service provider. So everyone is waiting in front of the starting line.

Agent payments are possible conceptually today, and you can experiment and build a business. But they fall well short when it comes to remedies for harm or fixing mistakes. With no centralized operator, the E-Commerce Act has a hard time protecting anyone, the Electronic Financial Transactions Act wasn't written for this ecosystem, and the Digital Signature Act only provides the most basic concepts. So anyone jumping in has to run a business, and take part as a consumer, without legal protection. Suhan expects that once the private sector gets active, the Financial Services Commission, which puts financial consumer protection above everything, will try to design rules from a centralized point of view, and it will take more time for regulation to catch up. On businesses that invest early and can't hold out: if you look at the finances of Korean virtual asset service providers from 2017 to 2022, apart from one or two exchanges, almost none grew enough to fund themselves while waiting four or five years for the rules to arrive with no revenue model.

Asked whether securities firms are preparing for the onchain era, the way Mirae Asset acquired Korbit (a Korean crypto exchange), Suhan said securities firms exist to give people value through investing in today's capital markets, so they aren't the ones researching new protocols at the frontier. Instead, the way BlackRock made a bitcoin ETF first, they were early to see virtual assets as an investable asset class, and in the US it has become possible to some extent to buy and sell stocks with stablecoins through exchanges. What securities firms mainly care about is tokenizing every asset that can be securitized, because tokenization makes identity verification easier and leaves a record of every trade.

When asked for one rule for anyone who wants to hand payments or trading to an AI agent, Suhan said to set your risk limit first. None of us are at the very front, so we need standards that someone recommends and helps with, but even within those there's a line you have to draw for yourself.

Set a limit you can actually take responsibility for. Figure out how much you can handle even if you get a result you didn't intend, and build that into a protocol or an algorithm.

How to remedy things that have already gone wrong is something the government and lawmakers will have to think hard about. What individuals can do right now is prevention. Next Securities is building its MTS (mobile trading app). The goal is a platform that's easy for people who want API trading, with customers who actively express their intent through AI agents as its core users.

t54 was answering the same question with guarantees and evidence, the protocols with limits and refund mechanisms, and the law with the principle of self-responsibility. Nobody yet has a single place where the answer to "who's responsible for money an agent spends" ends. I think whoever fills that gap first will set how fast agent payments make it down to everyday shopping.

Originally written in Korean on Brunch, September 28, 2026.