Nuvei：AI 替你付款，刷錯了誰負責？

8:07:36開場

8:07:36我上台可以嗎？在座各位真的很勇敢，這是今天最後一場 panel，辛苦了。希望沒有人在等後面的酒會吧？好，我們快速自我介紹。這是今天最後一場 panel，從我右手邊開始？好，從右手邊。我還以為你要反過來。沒關係。大家好，我是 Jackie，在 Bridge 負責 crypto 與發行相關資料。I'll come up. Is that okay? So I need to say everyone here is really brave. Because this is the last panel, so well done. I hope no one is waiting for the drinks for later, right? Okay, cool. Let's do a quick round of intro. This is our last panel of the day. Do we want to start on my right? On my right first, yes. I thought you were going to do, like, a reverse thing. Yeah, no, no, it's fine. Hi, everybody. My name is Jackie. I lead crypto and issuance data at Bridge.

8:08:05Bridge 是 Stripe 旗下公司，做 stablecoin 編排；Stripe 其實觸及 stablecoin 資金流動的每一跳。大家好，我是 Alice，Nuvei 數位資產副總裁。我們是支付處理商與收單機構，足跡遍及 200 多國，從 2017 年 crypto 剛起就支援 crypto 業務到現在。Bridge is a Stripe company where we do stablecoin orchestration, and Stripe really does touch every hop of the money movement of stablecoin. Hello, everyone. My name is Alice, and I'm Vice President, Head of Digital Assets at Nuvei. We are a payment processor and payment acquirer present in more than 200 countries, and we have been in crypto and supporting crypto business since the beginning, 2017 until today.

8:08:38大家好，我是 First Digital 的 Vincent Chok、執行長，最為人知的是發行與美元掛鉤的 stablecoin FDUSD，大約三年半前首次發行。謝謝各位，尤其是整個 TOKEN2049 週可能還留到最後一場 panel 的勇士們。Hello, everyone. Vincent Chok from First Digital, CEO of First Digital, most known for issuing FDUSD, a U.S. dollar-referenced stablecoin. We first issued that three years, three and a half years ago. Hey, everyone. Thank you. All the troopers that are staying back for maybe the last panel of the whole TOKEN2049 week.

8:09:05我是 Hassan，Coinbase 新加坡國家總監。我們在新加坡受 MAS 監管，服務一般消費者、合格投資人與機構；開發者相關的 Base 營運也在這裡，那是我們的 layer 2 區塊鏈。橫跨託管、交易、stablecoin 發行與流通，還有代理支付用的 x402 協議，等等會談到。I'm Hassan. I'm the country director for Coinbase in Singapore. In Singapore, we're licensed by the MAS. We provide services for consumers, accredited investors, institutions. We also have our Base operations here for developers, which is our layer two blockchain, and we sort of go across the spectrum of custody, trading, stablecoin issuance and distribution, and then also our x402 protocol for agentic, which I'm sure we'll get into.

8:09:31好，謝謝介紹。我是 Francesco，帶 MetaMask 的 developer relations。今天談 A2A、代理支付，當然也含 stablecoin。這是最後一場，我們會稍微辣一點，大概 30 分鐘。我想先開場：x402 支付量創新低，ERC-8004 的 attestation 也……Cool, sounds good. Thanks so much for the intro. I'm Francesco. I'm leading developer relations at MetaMask, and yeah, I'm excited to have this. So it's a mix of, like, A2A and agentic payments, stablecoin included, of course. I think this is the last session, so we'll try to be a little bit more spicy. We have, like, 30 minutes. One thing that I want to kind of, like, start with: x402 payments volume is at an all-time low, and ERC-8004 attestations are . .

8:09:59現況與大玩家

8:09:59大玩家也還沒真正採用。我想先問：你們怎麼看代理支付今天的狀態？可以同意也可以反駁，自由發揮。好，一開場就扔手榴彈。代理支付、代理商務、代理金融，大家都清楚還非常早期。大家對潛力、TAM 都很興奮，我們也都在往那走。not super adopted by also, like, bigger players. I just want to ask, what do you think is the state of, like, agentic payment today? Feel free, anybody can disagree, agree, and just talk. Sure, yeah, thanks for throwing that grenade right in there from the start. Yeah, look, I mean, I think we all— we all know agentic payments, agentic commerce, agentic finance is, you know, very, very early days. I think everyone's excited about the potential, you know, the TAM, and I think we're all sort of working towards that.

8:10:29會有很多實驗、很多有趣元素正在拼在一起。但有一點很清楚：開發者關注、資本、企業採用都夠多，這個領域很可能會變大。It's going to be, you know, a lot of experimentation, a lot of, you know, grounds of sort of interesting things that are coming together. But one thing is clear, that there is enough developer attention, enough capital. You know, enough enterprise adoption and so on happening in this space that, you know, it's probably going to be big.

8:10:46能長多大，想像力幾乎是上限。就現場動態而言，交易量有尖峰，從小基數算有穩步成長；目前 traction 多半在「準專業」開發者，為自己建工具、自行啟動生態。And I think sort of the imagination is the limit as to how much. What we see in terms of just, you know, on-the-ground action today, I think the volumes have been spiky. They have sort of grown relatively steadily, just— sort of off a small base, where most of the traction today seems to be around, like, you know, what I would call, like, sort of prosumers, developers that are effectively building for themselves to bootstrap the ecosystem.

8:11:13我們有叫 agent.market 的 discovery 入口，讓開發者與 API 供應方上架資源，代理與人類都能去發現。We have, we have, like, a discovery portal called agent.market, which is a way for, you know, developers and API issuers to really kind of, you know, put up resources on the supply side, and then for, you know, agents and humans to be able to go discover.

8:11:27上面大約有 2.3 萬筆資源。很令人期待，正在成形；我們也依使用者與開發者回饋擴功能。我們看得清楚這會變大，不必太擔心交易量尖峰的樣子。我同意，我是看資料的人，天天盯數字。There's about, like, 23,000 resources there right now. But it's, it's, you know, I think it's exciting. It's coming together. We're trying to build out a lot of functionality that we get, like, feedback, you know, from consumers and developers from. So it's something that we have, like, line of sight of: look, this is going to get bigger. But I wouldn't worry too much about, like, the spikiness of, you know, what volumes look like. Yeah, I was going to plus one to that. Like, I'm a data person, so I watch the data, like, day in and day out.

8:11:55若只看 X402 的絕對交易筆數：去年是這樣，後來變那樣。顯然去年某段「冬天」有人在測試之類的。And yeah, if you just, like, look at the absolute number of transactions for X402. Like, last year it was like this, right? And then it went like this. So clearly, like, something was happening last year in, like, the winter stage. People might be testing, whatnot.

8:12:10兩大市場與早期感

8:12:10但談代理支付，我會分兩塊市場，真實採用其實已經在發生；就算數字還小，也不代表「做得不好」。But I think when I think about agentic payments in the two, I would say, markets, I think we see the real adoption already. It's a matter of even if it's, like, I think, smaller numbers, it doesn't really mean it's like, oh, it's, like, not going well.

8:12:25例如機器對機器支付，X402 與 MCP 都在被採用，賣方或 server 數量穩定上升。人類參與較多的代理商務這邊，採用也很多。So, for example, on the more, like, machine-to-machine payment side, we see both X402 and MCP being adopted, and then you can see the number of, I guess, servers or the sellers actually steadily going up. And then on the side of where the agents are... involved in the more human-involved, like, agentic commerce side, you also see a lot of adoptions.

8:12:45新協議數量也誇張，昨天才勉強跟上 Machine Commerce Protocol，Sierra 又宣布什麼 Personal Commerce Protocol 之類，縮寫快被用光了。And then another data point on just, like, the number of protocols that are coming out is crazy, right? Just yesterday, I was barely keeping up on, like, oh, like, Machine Commerce Protocol, and then, like, Sierra announced this thing. It's, like, personal. I think Personal Commerce Protocol or something. All the acronyms are getting taken up faster. Yeah, yeah, yeah.

8:13:05監管明朗之後，明顯有巨大被壓抑的需求；現在是基礎建設在堆，接下來輪到台上各位想辦法。So you can clearly see that, like, there's a huge pent-up demand, like, after all the regulatory clarity. So I think right now, just, like, the infra is building, and then it's, like, kind of up to all of us on the stage to kind of figure out how.

8:13:19怎麼把真採用推進下一階？怎麼讓階躍函數往上？知道該把力氣花在哪、哪個 use case 很重要。代理支付很重要；我們幾個月前也推出 MetaMask agent wallet 當交易執行層。How do we bring this, like, real adoption into the next stage? Like, how do we see that step-size, like, function going up? I think it's super important also knowing where to put your calories in and understand which use case. I think agentic payments is super important. We launched a couple of, like, months ago also the MetaMask agent wallet for, like, trading, like, just an execution layer.

8:13:39進階交易者確實在上面跑策略。這兩大方向值得投入。請講。代理支付就像你在機器人競賽看到的，幾乎在嬰兒期，機器人跌跌撞撞。消費者對代理支付也是謹慎試水。Definitely, we see advanced traders basically executing strategies on top of that. So I think, like, these two main areas are worth, really, to see and grow and put our calories on. But go ahead. Yes. Well, I think agentic payments is really just like what you see. I mean, I was watching the robot games. It's almost, like, in its very infancy. You see these robots just stumbling everywhere, falling off left and right. I think that's the way that the consumer is also treating agentic payments, is let's cautiously see how it works.

8:14:12身為 stablecoin 發行者，我們當初做 stablecoin 根本沒想過代理支付，是為交易、DeFi 提供流動性，那條線仍在。I mean, for me, as a stablecoin issuer, I think that's, it's really a function where, when we first started a stablecoin, A stablecoin, we never even imagined the agentic payment side of it. It was, you know, provide liquidity for trading, DeFi, and that's still going on, right?

8:14:30穩定幣與四方可行

8:14:30這也會更多走進代理場景：人用 stablecoin 與 AI 代理 24/7 盯盤，凌晨 3:43 搶最好比特幣價。還不會用代理去買房結算，也許有一天。And I think that's also going to morph more into the agentic space, people using stablecoin and their AI agents to stay up all night, 24/7. To get the best price they can for a Bitcoin trade at 3:43 a.m. in the morning. So it's still, but it's not going to be where, you know, we're going to go and settle, you know, purchase a house using our agent at this time, but maybe one day.

8:14:57請講。從支付處理商角度，我們已上線讓代理 AI 互動的平台，從終端使用者到商家 AI；跟 Visa 網路與歐洲試點商家測過。代理 AI 潛力我同意，但極、極難，我們做過。Please, please. I can talk from— The payment processor point of view, because we actually launched the platform for agentic AI to interact with each other, from the end user to the merchant AI. So we tested it with Visa network and also with some pilot merchants in Europe. If you are talking about the adoption of agentic AI, yes, I agree that there's a— We have potential for everything. But it is extremely, extremely hard because we did it.

8:15:25真的很難。至少四方：終端使用者、替使用者發行代理 AI 的機構（不一定是本人）、中間商家、還有我們支付處理商。It is really, really hard. We need to have at least four parties involved. The first one is the end user. The second one is the end user's agentic AI issuer. Not necessarily them. They are using someone else to issue the AI to them. The third one is the merchant in between, and the fourth one is us, the payment processor.

8:15:45四方都要用同一套語言溝通，技術上很難。第二還要整個生態系，Visa、Mastercard、Amex 等都要認代理 AI 支付。So we have at least four core members, and all four of us need to talk in the same language as well. So that is really hard technically. And the second thing is, of course, there's the whole ecosystem around it that we need to have.

8:15:59上面還需要監管者定義每一步誰負什麼責。談採用，想法與夢都在，但要大規模上線並不容易。Visa, we need to have Mastercard, we need to have Amex, for example, to also recognize that agentic AI payment. And on top of that, we need to have a regulator who will define who is responsible for what in every single step. So if you are talking about adoption, yes, I believe the idea is there, the dream is there.

8:16:19請講。Bridge 也發 stablecoin，發行要採用本來就很難，Circle、Tether 擋在那，大家都想搶流動性，非常不容易。But in reality, in order to launch it for huge adoption is not so easy. Yes, go for it. I guess I just want to, because at Bridge, we do issue stablecoin also, and I think similar, maybe similar. I think it's an extremely hard market even in just, like, issuing. Stablecoin to get that adoption, right? We all know there's both, like, Circle and Tether, and, like, everybody is like, oh, we want to do the liquidity situation, but it's very hard.

8:16:50從發行到支付

8:16:50從 stablecoin 發行走向支付場景，我確實看到；所以 x402、MCP 今天多半集中在買基礎設施或買資料，我完全同意。So getting from more of stablecoin issuance to more payments land, I definitely see that, which is why I think today the adoption of both, like, x402 and MCP are more focused on, like, the buying of infra or buying of data side. And I completely agree.

8:17:06堆疊各層需要自己的協議互講。MetaMask、Coinbase Wallet 在入口層做得很好，那是第一關；有人提到信任問題。I think the different parts of the stack need to have, like, their protocol, basically, to talk to each other. And I think what MetaMask and Coinbase Wallet are doing extremely well is the first access. Like, the first point of this stack, which is the access layer, I think this—you were saying about the trust, somebody was saying the trust issue.

8:17:26我不會完全信任代理幫我買機票，雖然很多人已經這樣做；我不確定代理是否拿到最划算的方案。Like, for me, like, I'm not gonna trust my agents to just, like, buy me even, I don't know, even air tickets. I know a lot of people do that already, but, like, for me, I don't know if, like, my agents are getting the best deal for me, right?

8:17:38若把銀行權限交給代理，有人說存貸能賺更好收益，我還沒信任到那步，需要入口授權讓我看到真能信任代理支付，核心或企業級 use case 的採用會晚一點。And I don't know if, like, if I give— My agents, like, banking access. There's a lot of talk about, like, oh, you can maybe get better yield for your bank saving nowadays. Like, I don't have the trust in my agent yet, so I need the authorization of the point of, like, the first point of entry to show me that I can really, like, really trust the agentic payment technology for it to succeed. So I think the adoption of, like, more core use cases, or, like, including the enterprise side of payments, will come later.

8:18:08我們上線平台與試點後也談信任。總結要兩件事：把代理 AI 關在規則框裡。I can add on to that because we actually launched the platform, and then the pilot, right? When you're talking about trust, I totally agree, and we also come to the conclusion of how to help this adoption. In summary, there are two things that need to be done. First one is you need to be able to put your agentic AI into a box of rules.

8:18:30給你 10 美元上限就待在 10 美元內；限定品類就別越界。AI 的 intent 必須守這個框。I give you a limit of $10, you stay in that limit of $10. I give you this category of items, you stay in that category of items. So it's the box of rules that your AI needs to have, that intent to stay in that box of rules.

8:18:45第二，支付尤其如此：你不信任代理代付，但若在我這買可以反悔、撤銷交易，你就敢用。研究與實際上線的結論是：規則框加上能把錢拿回，才會採用。And the second thing is, for the payment especially, is you don't trust your agent, right, to do the payment. But if I say to you, you can use your agent to pay to buy a product with me, and you can change your mind. You can reverse the transaction, then you would use it, right? So that's the second thing that we found from our research and our actual launch of the agentic payment platform is if we give the end user the box of rules for the AI to operate and the flexibility in taking my money back, then you can have the adoption of that technology.

8:19:19護欄與機構

8:19:19Alice 說的我非常同意：給代理專用護欄、沙盒化行為，還有 delegation 流程。現場舉手：誰願意把 100% 財產交給代理？100%？看到了吧。這是信任問題，但我們也在打地基，這對產業長大最重要。Yeah, I second that a lot, Alice. Basically what you're speaking about is giving guardrails specific to agents, sandbox their actions, and also delegation-specific, like, processes. So I think that's easier. But maybe let's check it out. How many of you guys here, raise your hand if you will give 100% of your wealth to your agent? Huh? 100%. Do you see? Yeah, I mean, I think it's definitely, like, a trust, right? But also, like, what we're doing here is building fundamentals. I think that's the most important part to set up, like, literally an industry that is flourishing.

8:19:58我們談過零售。機構採用代理，各位看法是否不同？可能有點難答。機構現在有合規、監管、申報義務，還有 CRS、FATCA。We kind of, like, spoke on retail. Like, do you guys see a different kind of, like, take around maybe institutional adoption of agents before I'm saying what I'm thinking about? Yeah. Might be a little difficult, sorry. I think for institutional adoption of agents right now, I think it's because institutions are, you know, we have compliance issues.

8:20:19理想世界是代理幫你搞定一切，甚至彈指拿到 MAS 牌照，但不可能，現場也沒監管官。很多 use case 其實沙盒內的自動化，不算代理 use case。We have regulations, we have reporting obligations. There's also now with CRS and FATCA. It's, it's, yeah, it'd be nice to live in that world where your agents can do everything for you, including get you an MAS license in a snap of a finger.

8:20:39但企業這邊我看到的切入點是：企業 use case 太多，財務功能尤其熱，例如應付帳款、庫藏再平衡；代理到某種量產規模後，很可能配專屬錢包或授權，真正開始花錢、搬錢。But, you know, that's impossible. No regulators here. I think there are a lot of use cases that are really, like, sandboxed inside, so they are not even agent use cases, right? But they are automation. Yeah, I think for enterprises, the way I look at the attack factor here is that there's just so many, you know, enterprise use cases, you know, where . . there is a lot of energy around, like, you know, applying agentic AI. And I think especially, like, in the finance function, you know, around things like, you know, vendor payables and treasury rebalancing and so on, that's most likely going to be the place where, you know, once you have maybe agents at some level of production scale, you will start, like, arming them with their own wallets or their own

8:21:19授權去實際支出、移動資金。今天我看到的多半是自動再平衡、加快速度，還沒真正放出有 intent、有完整脈絡與完整授權的代理。authorizations to actually start spending and actually sort of moving, you know, money. Today, it's, what I see at least today is, like, it's more about, like, automated rebalancing and, you know, making that faster. We haven't truly sort of unleashed agents yet, which have intent, which have full context and have full authorization.

8:21:37零售信任要件

8:21:37Vincent 說的，企業要大規模邁下一步，還得先理清治理、營運、財務風險。好。談代理專用護欄或安全，各位要補什麼？除了限額、一次性虛擬卡、代使用者 delegation，還缺什麼定義才能贏零售信任？As Vincent was saying, there's a lot of governance, you know, operational matters, financial, you know, risk matters that need to be sorted out before I think enterprises at scale will be ready to actually take that next step. Sounds good. Anyone want to add? It's good. Okay, I think, like, basically, like, speaking about guardrails or, like, security specific for agents, do you guys want to kind of, like, build on top a bit? Kind of like you mentioned a little bit about guardrails and what type of kind of, like, delegation you want the agent to have, but what are kind of, like, the minimum things that we need to kind of approach, especially for, like, getting more retail trust?

8:22:19是限額、支出護欄？一次性虛擬卡我也看過。還有代使用者帳戶操作。還需要定義什麼？Is it like... You know, guardrails on limits, on spend. I see, obviously, virtual cards issued with, like, one-shot spend. What else? I see delegation on a user account on behalf. What else you think are kind of, like, something that we need to still define, yeah?

8:22:37我個人看到兩塊要補，第一很無聊：稅務與申報。做 crypto 的朋友多，聚餐結帳我問「有 Venmo 嗎？」美國人說沒有，「那 stablecoin 付我？」我得想稅務影響。I think personally, for me, I see two areas for improvement. One is actually really boring, on basically tax and reporting. I'll use a funny anecdote. Working, having been working in crypto, I fortunately have a lot of international friends. But when we go out together, we need to settle. Like, let's say we go out to dinner, I'm like, "Hello, do—" "Have Venmo?" They're like, "No, U.S. people." And they're like, "Why don't you just, like, pay me in stablecoin?" And I'm like, "Well, then I need to worry about the tax implication of this, right?"

8:23:10某些服務我能用，以前還幫客戶寫查詢做 crypto 支付申報。個人層級的稅務申報能力還不夠。發行 stablecoin 做資料，每個月 attestation 申報，都還不算全自動，代理那天會到，但稅務、會計、申報還要很多改善。Like, I know that I can, certain service, I can use it, and, like, I literally used to write queries for customers to help them with, like, reporting of their crypto payments. But I— Like, this is just a small example. Even on personal level, tax reporting, like, capability is just not there. So having, like, been in a data function for issuing stablecoin, I see, like, every month we need to do attestation reporting, everything.

8:23:33第二塊是授權，留給其他與談人細談。這兩邊都要更好，才有真採用。It's not, like, even super automated yet. So I don't even know, like, you know, I know we'll get there one day with agents. So I think on the tax, accounting, reporting side, there's a lot of improvement. And then the second one is on the authorization, which I will leave to the panelists to talk more.

8:23:50教育與 Web2

8:23:50對我來說這兩邊都要進步。採用現在盯零售、簡單支付、小額、零碎支付，但零售有沒有真正教育、訓練、資訊教怎麼用代理支付？But I think these two sides need to improve more, for me too, for there to be, like, real adoption. To me, I think it's gonna— this is kind of a double-edged sword. Because I think the adoption is really right now is to look at the retail consumers, very simple payments, fractional payments, small payments.

8:24:11很多機構自己資產負債表、客戶帳都還管得草率。若給零售一個錢包說愛存多少存多少，有人就把 100% 存進去因為很酷。However, those consumers are also— is there any real education, training, information on how to use agentic payments? You're telling there's a lot of institutions that really are still, in some ways, irresponsible in how they manage their own balance sheets and client balance sheets.

8:24:32若給零售一個錢包說愛存多少存多少，有人就把全部資金存進去，只因為覺得很酷。出事時因教育不足、外界資訊不清、代理釣魚代理或詐騙而賠錢。So if you're giving a retail person themselves, well, here's a wallet. Deposit as much as you want. They decide, "Hey, I'm going to deposit 100% of my money into this because it's neat." When something happens, because of the lack of education, a lack of clarity in what's out there, agents phishing other agents, scams, they lose their money.

8:24:58他們怪誰？錢包開發者？stablecoin 發行者？問題還很多。這是消費者議題，監管存在就是保護零售。我們都在機構端，自己也還在學；要大規模採用，教育很關鍵。Who do they blame? Do they go after the creator, the tech of the wallet? Do they go after the stablecoin issuers? Do they go after— So there's still a lot of question there. And this is the consumer. This is why regulators exist, is to protect the retail consumer. So with that, I think, you know, us here all from institutions, and we're still learning this stuff. But to get this to mass adoption, I think it's really about education.

8:25:25Vincent，你提的開放問題也包括代理 attestation 基礎與誰負責。在完整 attestation 生態成形前，X402 是很大一步。我的延伸問題是：Web2 大型基礎設施商似乎還沒真正對齊這些標準，對吧？Yeah, I think also, Vincent, like, one of your open questions is also about building the fundamentals around attestation for agents and who is liable for that. But until we have a proper, like, attestation ecosystem, and I think X402 is a big step forward. My kind of follow-up question or segue is, we don't see also, like, Web2 big infrastructure providers really, like, aligning with the standards, right?

8:25:54大家各建各的信譽系統。要問各位：怎麼跟更大 Web2 玩家對齊？若 Web3 軌道強在鏈上 attestation、可追溯與信任，卻沒帶進來，或標準很好但只有市場上極少數代理在用，算成功嗎？Because everybody is kind of, like, building their reputation system, right? And maybe, like, a question to everyone here is that, you know, how we can do a better job also to align with, like, bigger Web2 players. Because if we're actually, like, not either, you know, bringing some of Web3 rails, and why Web3 rails are strong is because, like, you can have on-chain attestation, you can have proper, like, traceability, trust, right?

8:26:19責任與信譽

8:26:19對 Web2 應用其實很合適，但怎麼更橋接兩邊？若標準很好，卻只有市場上 0.01% 代理在用，很難算成功。看 X402 Scan，六個月前那些代理幾乎還是同一批。And I think it's a perfect kind of, like, fit for Web2-specific applications. But how we can kind of, like, bridge the two more, right? Because I don't think, like, we can define success if, you know, we will have a great standard, but those great standards are used, I don't know, by 0.01% of actually the agents on the market, right? That's right. If you look at X402 Scan, we have literally the same agents that we had six months ago.

8:26:46大概只多 3% 到 4%。我們剛剛也繞到責任：責任怎麼分，才能進一步建立信任、把生態做大。I think it's, like, plus three or four percent more. Yeah. So I think one of the maybe topics that we nibbled around for which you're asking is, you know, about this topic of liability and, you know, how that . . basically can unlock further trust, you know, in order to actually start sort of growing out the ecosystem.

8:27:05今天 attestation 還更高一層，像代理市場的信譽系統，有點 eBay 模式。金融服務底層需要更強，需要終極責任。卡組織兩端發卡行、收單行承擔消費者或商家風險、管理組合，就是這個邏輯。Because today, I mean, even attestations is one level above, right? It's a reputation system that's frankly like a marketplace of agents, and you sort of go off of that in sort of an eBay model. But for financial services, you know, underpinning, you need something stronger, and you need, like, ultimate liability. And so that's how, like, if you look at card networks, you know, where you have issuers and acquirers on either side, they're the ones who are, you know, underwriting the consumer or the merchant, and they're sort of assuming that responsibility and then managing a portfolio.

8:27:34新加坡 MAS 有創新中心，其中 Project Bloom 我們參與代理端，團隊核心工作之一就是釐清責任歸屬。Locally, like, the MAS does have this innovation, you know, hub, and one of the projects they're running is called Project Bloom. We're participating in that on the agent side, and this is really one of the key things that that team is trying to sort through.

8:27:48責任最終歸誰？可用 delegate 模型：做完 KYC 的個人或 KYB 機構，代理是其 delegate，責任仍留在背後的人或機構。Okay, how do we think about liability? You know, who does that ultimately belong to? You could even sort of use a delegate model and say, look, this is the KYC sort of individual or, you know, KYB-like institution or business, and then this, this, you know, agent is their delegate, and so you retain liability.

8:28:06但代理再 spawn 子代理，加上 stablecoin 與嵌入式錢包，分岔擴散太容易。我們才剛觸及第一層 robust 模型該長什麼樣。But then what happens when, like, agents start spawning sub-agents, and then especially when you layer on stablecoins and embedded wallets, and just it's very trivial to kind of fan out from there. So, so I think we can. We're starting to scratch the surface on what a robust model will look like, at least on that level one.

8:28:25在可組合、可程式、stablecoin 可細到百萬分之一美分的 frontier，會亂一陣子。我想稍微反駁 Web3/Web2 對立：問題不是 Web2 不採用，或 Web3 做得不好。But it's, it's, you know, just given we're in this brave frontier where everything's composable and programmable and, you know, you can take stablecoins down to a millionth of a cent, it's going to be, it's going to be whack for a while. I actually want to push back a little bit on this Web3/Web2 divide, because I don't think, like, it's the problem, like Web2 not adopting or that Web3 is not doing a good job.

8:28:49共同設計介面

8:28:49是我們有新的設計介面，要一起解。談 stablecoin 代理支付與區塊鏈本質，有結算問題，也有交易對手風險；Web2 法幣軌或 stablecoin 軌都要處理。I think it's more on we now have new design interface, and then we need to tackle it together. So I think when we think about, like, stablecoin agent payments, including the nature of blockchain, there's the issue of settlement, and there's also the issue of counterparty risk.

8:29:08結算這邊，鏈上 stablecoin 交易「quote-unquote 結算」了，仍要做對帳；即時結算表面下，交易量與交易對手更多，對帳面更大。And both, regardless, like, are you in the Web2 world? Are you in, like, fiat rail? Are you on, like, stablecoin rail? Like, they are problems that we need to deal with. So. So on the settlement, even though on the blockchain, for example, the stablecoin transactions, they quote-unquote settle, but they don't actually settle because you need to still do reconciliation.

8:29:29產業要解這個。交易對手風險則是共同語言，所以我雖然開玩笑協議縮寫爆量，協議其實很重要。代理要互相確認身份；若是 JPMorgan 代理我可能給 key，隨機 marketplace 代理我不信。It's like you have instant settlement, but actually the surface for you to do reconciliation is just, like, much more because you have, like, much more transaction volume, and there's, like— way more counterparties you need to deal with. So I think as an industry, we need to figure that out. And then on the counterparty risk, it's like having the shared language, which is why, as much as I joke about, like, all the, like, acronyms, the protocols are actually very important.

8:29:57這是很 exciting 的時刻，Web2、Web3 正在合流成要一起處理的原則。各位有在做 know-your-agent、代理 KYC 嗎？The agents need to acknowledge to each other who you are. It's not that I don't really trust, like, an agent. If I know you're a JPMorgan agent, I will probably give you my key. But, like, I'm not gonna trust a random agent on, like, this marketplace. Yeah, so I think it's actually a very exciting moment where we're co-designing the whole Web2, Web3 thing.

8:30:19要 unpack 的太多，我試著跟剛才主題。跟 Web3 比，代理支付其實是機構在推：Visa、Mastercard、我們 Nuvei 這類 PSP 是先行者，套用既有支付技術，不是純開發者發 stablecoin、比特幣。I think it's, like, merging into principles that we need to, like, tackle together. Are you guys, like, looking at know-your-agent kind of, like, actions or, you know, KYC for agents? Is this something that is happening, you think? Wow, there's so many things to unpack, so I try to follow the topic that you have just mentioned. Let's talk about comparison between this new technology to Web3, right? What I found the difference is agentic payment is actually driven by institutions, right? So Visa, Mastercard, and us, like Nuvei, we are payment processors.

8:30:54至少我們從政策與監管理解基礎設施。代理支付採用很難，但立法與政策輪廓我們已知道。We are actually the pioneer in applying the technology, the payment technology. It's not something that— we are developers, and then we are building some, like, stablecoins, or we're building some Bitcoin or something like that. So at least we understand the infrastructure from the policy and regulation point of view. So if you're talking about adoption for agentic payment, yet it is very hard, but we already know the legislations and the policy behind the scenes.

8:31:19PSP 與 KYA

8:31:19理論上的基礎設施已在，現在要實際放進網路與技術。Hassan 提的責任，不幸很多會落在 PSP 身上。So the— infrastructure from that theory is already there. Now we need to practically put it in, like, network or technology. The second thing is about the liability that Hassan mentioned a little bit. The liability is unfortunately going to be falling onto the PSP a lot.

8:31:40我們建平台讓使用者代理跟商家代理對話，但我們解碼訊息與 intent。若使用者代理 hallucinate 買錯、我們沒抓到，我們要負責。For example, we build the platform and we allow the end user agent to talk to the merchant agent, but we are the one who decode the message and the intent. So if your end user agent is hallucinated and make the wrong decision on the purchase, and we couldn't detect it, we are liable for it.

8:31:59KYA（know your agent）也是我們做，責任從商家轉到 PSP。我們是否願意為交易負責？這是 PSP 採用的巨大問題，跟終端、商家節奏也不一致。And we are the one who doing the KYA, the know your agent. So it's the whole liability shift from the merchant to the PSP. So are we comfortable in being liable for the transaction? That's also a very big question. So the challenge is the adoption from the PSP point of view, not at the same time with the end user and the merchant point of view as well.

8:32:29合規上我們做 know your agents，本質仍一樣：誰控制代理與錢包？一人可開上千代理，真正要 KYC 的是那個人與錢包。As for us, when we conduct any compliance for know your agents, but we generally still do the same thing. Who is the controller of the agent and the wallet itself? You know, I think it's impossible to say, how do you conduct KYA on the actual agent when one person can have multiple agents, thousands of agents, but that one person is the person that you really need to do the KYC on and the wallets.

8:32:59只是多一層，有時我們把問題複雜化。託管、受託人角色我們做 KYC 很多年了，先簡化，等工具一鍵連點再說。So, I mean, it's just another layer. I think that sometimes we just overcomplicate things. So for us, we, as, you know, our custodian and trustee hat, we've been conducting KYC for many years, and so that's the way we look at it. We just try to keep it very simplified until— there's such tools available that it's just a snap of a finger and it connects the dots.

8:33:28好，剩約四分鐘，現場有問題嗎？有。PSP 說代理對代理交易責任在 PSP，是自願承擔，還是監管框架規定？對，這正是我們問自己、也問監管者的：整個責任轉移。Sounds good. I think we have like four minutes left. Does anybody have any question? Nothing much? Oh yeah, nice. So, PSPs, you said that the liability falls on the PSPs when agents transact with agents. Is it the PSPs taking on, voluntarily taking on this liability, or is there some part of a regulatory framework that says PSPs got to be liable for this? Yes. Well, that's the exact question that we are asking ourselves and also asking the regulator, because it is a whole liability shift.

8:34:11傳統爭議與自上而下

8:34:11傳統世界詐欺、爭議、chargeback 會推回商家：你 3DS 沒做好、風控沒開等。In the traditional world right now, if there is any fraud, dispute, or chargeback, we push it back to the merchant, and we're saying that because you did— didn't do a good, like, 3DS, or you didn't turn on the security or risk control, right?

8:34:25現在代理 AI 是我們解碼訊息，讓使用者 AI 跟商店互動；我們是平台、中間人，要負責，這是巨大系統更新。But now with the agent AI, we are the ones decoding the message to allow the AI from the end user to interact with the shop. So we are the platform, we're the middleman. We are liable for that, and it's a huge... system update.

8:34:40需要新法遵、新律師、新合規團隊，盡職調查對象是代理而非只使用者或商家。等監管者給完整政策與步驟：我哪些要負責？爭議流程也要重做。會很久。So we need to have a new regulator, new lawyer in the company, new compliance team, due diligence of the agent, not the end user, and not from the merchant. And we are waiting for the regulator to give us the whole policy and step by step, like what do I have to be liable for, right? And then it's going to be the whole dispute process as well.

8:35:05相對 Web3 先建再碰監管，代理支付是自上而下：知道痛點、從挑戰往下做實務，我認為路線對。So it is going to take a long time. But what I'm trying to say is, compared to Web3, we know the pain point, we know the challenge, and we are tackling from the challenge down to the practical. Versus building, building, building, and then we are tackling the regulators later, right?

8:35:26若 PSP 要負責，代理擁有者不負責嗎？你們不調查嗎？看錯誤類型，關鍵是 intent：能否解碼 AI intent 是否與使用者設定一致，一切都有 code。So for agentic payment, it's actually going from the top down. And I think it's a good, it's the right pathway to go. So, sorry, I'm also curious because if you guys are going to be liable, does— the owner of the agent not liable? Do you guys not do any investigation? Depends on the errors happened. It's the intent. So that's why we have to be able to decode the intent of the AI, whether that intent is exactly as what the user put in.

8:35:56intent 來自使用者沒錯；若 AI hallucinate 做錯決定，錯在 AI 不在使用者。但我們解讀訊息時沒對照使用者 intent，就我們負責。Everything is coded. But the intent is from the— user, isn't it? Yes, but if I'm just talking in the way that if the AI is hallucinated and then it made the wrong decision, the decision wrong because of the AI, not of the end user.

8:36:10若終端使用者是壞人，intent 看得出來。若 intent 就是買 100 美元、AI 照做，就不能回來 chargeback。但很多人 prompt 付款的用語可能出錯。But then we interpret the message, and then we didn't compare with the intent of the end user. So then we are liable for that. If it's the actual... The end user is the bad actor. If the end user is actually the bad actor, we can already see it from the intent, right?

8:36:28意圖與穩定幣試點

8:36:28換句話說，若 intent 就是買 100 美元、AI 照做，就不能回來 chargeback。但很多人下 prompt 付款的用語可能出錯。But say, look, it is your intent. You want to buy $100, and the AI did exactly that, then you cannot come back to me and claim chargeback for that. But to prompt that payment, the way that users, a lot of them, basically their vocabulary, the way that they prompt these AI agents to make a payment, there could be a mistake there.

8:36:49若那種 prompt 失誤，你們也扛？不是，若我們依使用者寫進去的 intent code。所以要等監管者寫清楚政策：哪些算我們責任、哪些算使用者。And if they make that mistake, you'll take liability for that? No, if it's that mistake, because we follow by the intent, the actual, like, code that they put in. That's why— in order to answer to it, we are waiting for the regulator to write a policy and a clear-book guideline in, like, what is it that's falling under our liability and what is falling under the end user.

8:37:11那跟 stablecoin 發行者怎麼合作？付款已完成能召回嗎？很有趣：我們試點完全沒用 stablecoin。stablecoin 的 downside 是不可逆，我們代理支付不用 stablecoin。So how do you work with the stablecoin issuer then? Because if that payment's already done, do you recall that? That is really interesting. So I'm telling you, the pilot that we ran, we didn't use stablecoin at all. Okay. And that is the downside of using stablecoin. We don't use stablecoin for agentic payment.

8:37:31我也不建議 PSP 用 stablecoin，因為不可逆。試點用的是 tokenized 卡資料。若中間有一層，底層可卡可 stablecoin，上面做爭議等功能，就能用嗎？I see. And we don't recommend any PSP to use stablecoin because it is irreversible. What we are using—what we... used in our pilot, we used tokenized card information. And do you think there will be, like, a layer? Maybe if there's a layer on top where the rails are, you know, can be both cards and stablecoin, and then this in between can have the, you know, dispute, all the features, then you will be able to use it.

8:37:59對，底層卡加 stablecoin，中間可能是另一種 payment token。Jackie 在橋這件事。但 tokenized 卡與 stablecoin 性質不同：stablecoin 不可逆；tokenized 卡仍像傳統金融、仍視為貨幣。So, stablecoin. Yeah, so it would be like cards, stablecoin on the bottom, and then here, for example, it's another payment token situation. I think Jackie is trying to bridge that. But the character is different. The character of the tokenized card and the character of stablecoin are different. Stablecoin is non-reversible. For tokenized card, they are still like traditional finance, and they are still considered money, so it's different.

8:38:27我們沒用 stablecoin。我覺得 X402 適合信用卡額度以下的微支付，例如不到 10 美分用卡不合理，才用 stablecoin。The character is different. We didn't use stablecoin, and I think the only use case where, like, X402 can be used is for, like, micropayment outside of the credit limit. Let's say if it's less than— 10 cents, that doesn't make sense to use a card, then yes, you can use the stablecoin.

8:38:46爭議與貨幣定義

8:38:46但也缺 KYC 等，責任很大。代理支付我們目前不用 stablecoin。有人說 stablecoin 交易可以 reversed？那會跟整週 Web3 論述衝突。這週談代理支付與 stablecoin，Arthur Hayes 的 flop chain 不管名稱如何，有一點很重要。However, it doesn't have any KYC and everything as well, so it's a lot of liability there. Yeah, agentic payment. Unfortunately, we don't use stablecoin. Well, you know what? Stablecoin transactions could be reversed. Well, then you'd go against the whole Web3 industry. No, no, but I think this whole week, like, we are here talking about, like, agentic payments, stablecoins.

8:39:08機器對機器、次美分交易，或代理幫人研究的代理商務，stablecoin 要到你說的真採用，得先想清楚爭議怎麼辦；監管與產業一起定責任歸屬。I really think, who was it, like Arthur Hayes announced this, like, flop chain that he has. But I think he, regardless of the chain, the name, whatever, I think he has an important point. For whether it's machine-to-machine payments, the sub-cent transactions, where it's like agentic commerce, where, you know, an agent helps a human do the research, I feel like for the stablecoin to have the real adoption that you were speaking about, like, we really need to figure out how do we do the dispute situation.

8:39:39第一要知道交易對手是誰；第二代理流程出錯時，誰賠這個 mistake。時間快到了，可以聊好幾小時。And to your point, it's like the regulators, like together, like with industry participants, we need to figure out where the liability is. It's like, yes. So first of all, we need to figure out the counterparty, like you need to know who you are.

8:39:54第二，代理流程出錯時，誰為這個 mistake 買單，必須說得清楚。時間快到了，這題其實可以聊好幾小時。And second of all, we need to be able to answer if something got messed up in the agentic flow. Who is responsible to pay for that, like, mistake? Well, I know that we're running out of time. Yeah, we could go over, like, for hours.

8:40:08若要用 stablecoin 支付，得先確認該地是否認可 stablecoin 當支付工具、是否當「貨幣」。我早場 panel 有人說 stablecoin 不符合貨幣的經濟特性，我傾向不同意。If you want to use stablecoin for payment, you have to know whether stablecoin is recognized to be used for payment first, whether stablecoin is recognized as money to pay, and for— like, I was in the panel, like, earlier: stablecoin doesn't have the characteristics, or characteristics of economics, of money because it is— I tend to disagree.

8:40:31當然，若你有不同看法可以教我。有人說它「不會維持同一性」；除非 FDUSD 在每條鏈上都一樣，對，所以叫 1:1，我們有 attestation，每 1 FDUSD 背後 1 美元。有抵押，我不理解為什麼 stablecoin 不算錢。I'm sorry, yes, of course. If you have some other idea, you can maybe educate me. But at first, it doesn't stay the— same. Unless FDUSD stays the same in every single network, then yes, it has the— yes, that's why it's called one-to-one. It's, yeah, we have attestation. Every one FDUSD is backed by one US dollar. So you guys, it's backed, so it's not the— I don't understand why stablecoin is not money, like.

8:40:59結語

8:40:59新法規下 stablecoin 正走向現金等價。你說的是不可逆，那是 feature 不是 bug；對 PSP 來說，像卡組織那套日常運作確實難受，但也可以設計可逆性。有人說要看消費者怎麼想，例如 stablecoin 發行者能否 freeze 資金。Stablecoin is on its way to being cash equivalent anyway, with all of these new regs that are coming up. I think what you're getting at is it doesn't have reversibility, and that's a feature, not a bug. I get it makes the PSP's life difficult, the way that, like, you know, card network transactions run today. But it does have that. It does have reversibility. Yeah, I was going to say, it all depends. I mean, it's all the consumers. I mean, some consumers are saying, well, if— The stablecoin issuer has that kind of power to freeze money.

8:41:22不是制裁合規那種 freeze。她指的是卡有 180 天 chargeback 窗口，那不是 stablecoin 的本意，但可以圍繞它設計。Stripe 有 payment token，合併法幣、卡、stablecoin 軌；PSP 被付費是因為我們承擔風險，多扛風險也該被獎勵。No, no, it's not about, like, freezing for, like, sanctions compliance, etc. I think it's about, like, what she's saying is, like, you get a 180-day chargeback window with cards, right? Which, of course, like, that's just not, like, that's not the point of stablecoins, but you could design it around that. Yeah, I was going to say, at Stripe, we have, like, Stripe payment tokens, and it actually does combine the fiat rails, the card rails, the stablecoin rails, and the reason people would pay for the payment service providers is because we take on the risk.

8:41:52對，但該被獎勵。這 panel 越來越像在說服 Wayne stablecoin 是錢，該收尾了。我不是反對，是要在任何國家採用，得先被認定是貨幣。So we should be rewarded as payment service providers to take on the extra risk too. Yes, but you should be rewarded. I feel like this became less agentic and more convincing Wayne that stablecoin is money, so maybe we should wrap. No, no, I'm not saying that we are against it. I'm just saying that in order for it to be adopted in any country or regulatory region, it had to be first considered as money.

8:42:10不是說我們不當它用，但主題是代理支付：得先能在各國當支付工具，我們才能再往上加一層。I'm not saying that we don't use it as money, but because the topic is about agentic payment, right? So it had to be used for payment first, and it had to be recognized as a tool for payment first in every single country in order for us to put another layer back on top of it.

8:42:25Wayne，明年 panel 要接續這題。對，to be continued；明年有更多 use case 與試點可分享，今天試點用的是 tokenized 卡、Visa 網路，純支付場景。Wayne, I think you need to set this panel up for next year to continue. Yeah, I think we need to redo, to be continued. Yeah, next year we have more use cases and pilots to share, but I'm just sharing, like, what we are using today, the pilot that we used.

8:42:39完全沒用 stablecoin。好，可以散場了，我會發推文。是可逆是 bug 不是 feature 之類（笑）。panel 目標達成：大家都還醒著吧？最後一場，辛苦了，給與談人熱烈掌聲，謝謝。We didn't use stablecoin at all. We used tokenized card, and because it's only for payment, so tokenized card with the Visa network. Sounds good. I think we're good to go. I will definitely tweet it out. It's a bug. Wait, it's not a feature, it's a bug or something, yeah. And I think we reached our, you know, we reached what the goal of this panel is. Everyone is awake, right? And this is the last panel. So, well done, everyone. Big round of applause for our panelists. Thank you.