# Kevin Leffew, who joined Coinbase from Replit two years ago, pitches x402 for agent payments

Incline, a Stabledash show. Sep 14, 2026. Kevin Leffew.
https://live.stabledash.com/shows/incline/episodes/x402-explained-the-payment-layer-for-ai-agents-5xcxQZj1jFA

The x402 co-author says wallet signatures should replace API keys so AI agents can find, price and pay for tools on their own, with trading as the first big use case.

Kevin Leffew joined Coinbase about two years ago from Replit, where he was a go-to-market lead and helped launch Replit Agent. On Stabledash Live he made the case for x402, a payments protocol he says was built to solve three problems: tool usage for language models, economic reasoning for agents, and autonomous runtimes.

## Killing the API key

Kevin Leffew described the goal of x402 as to "kill the API key", replacing it with a signature from the client-side wallet that authorizes payment at request time. Today, he said, an agent with access to 700 plus tools needs 700 plus API keys, each requiring a human to visit a dashboard, enter a credit card and configure the tool call.

Under x402, an agent receives a 402 status code response carrying pricing information, so it can think economically about its tool calls. Leffew noted the 402 code dates to 1997, and argued that while humans find small payments like 30¢ transactions a cognitive burden, agents can execute micropayments at scale.

## Money as a reward function

Kevin Leffew called money a strong alignment signal for agentic workflows, saying an agent aligned with a user is generally aligned with that user economically. He said a wallet lets agents both earn and spend in a single runtime, without billing in arrears, credit or ACH transfers.

## An open standard

Kevin Leffew said x402 needs open source code and open governance, which led to its move to the X402 Foundation under Linux, with an open community at x402.org. He compared the path to Kubernetes, which Google incubated before moving it to the Linux Foundation.

He said inference is the dominant use case today, followed by tool calls, and that companies such as Browserbase, Exa, Browseruse, Apify and massive.com now support x402 and accept stablecoins.

## Trading agents with limits

In a demo, Kevin Leffew used Claude Code with Coinbase Payments MCP to pull top stock gainers from massive.com, finding ChargePoint up 75 percent. He said agents should underweight social sentiment and overweight official sources like PR Newswire, and that users should stay in the loop to verify trades at first. A wallet funded with $10 can spend at most $10, which he said limits collateral damage.

Kevin Leffew said Coinbase has no advantage on x402 itself, but does on trading, where paying for tools at request time matters. He expects trading agents to move toward autonomous execution over time, and sees new uses such as tip-steered video streams built on models like the one from Fall.

## Key takeaways

1. Kevin Leffew says x402 aims to replace API keys with client-side wallet signatures that let agents discover, reason about and pay for tools at request time.
2. x402 has moved to the X402 Foundation under the Linux Foundation, with partners including Cloudflare, AWS, Google, Stripe, Visa and Mastercard.
3. Kevin Leffew sees trading as the area where Coinbase has a structural advantage, and showed an agent that found a stock up 75 percent.

## Transcript

**Zach Fowler** (0:06)

Bringing the energy up. Last guest of the day. Saved the best for last. Arguably, maybe not. I don't know. We'll see. It's been it's it's been an AI master class today. It's been quite a day. We have Kevin, go to market lead, coauthor of x four zero two. Kevin, it's great to see you.

**Kevin Leffew** (0:25)

Thank you for taking the time today. How's it going? Hey. It's going pretty well. Thank you so much for having me on. It's always an exciting week in the x four zero two sphere around AgenTic payments and AgenTic trading. So excited to chat with you guys, answer any questions, and and go deep into to a demo later. Let's go. You're in the Bay Area. Right? Yeah. So I'm based in San Francisco. Nice. Work out of the San Francisco office at Coinbase, then we work right now in DC.

And I'll be in DC at StableCon next week Nice. Is around.

**Zach Fowler** (0:56)

Nice. It'll be a fun week in DC. We I don't know if you I know you're in the green room with Halsey's from MoonPay. He was giving some love to the x four zero two team, and it sounds like it's it's been a really key unlock for some of the the stuff that they're doing with Paybox and getting enterprises and developers actually going with some of this stuff, like X four zero two as an open neutral standard has been extremely helpful.

Maybe we just kick off with how how to think about x four zero two a little bit, like introduce our guests to what it is. This is the first time we've really dug into that standard and what it is. So we can start high level and then dive into some, like, specific launches and and other things.

**Drew Rogers** (1:40)

When did when did this start? Like, when did you start thinking, like, I gotta coauthor x four zero two. We we need to start this foundation. Like, you know, bring us into the give us the origin story.

**Zach Fowler** (1:51)

Yeah. Definitely. Maxi.

**Kevin Leffew** (1:53)

Sure. We want that. So much so much to talk about. I can kinda go back to the the origin. So I joined Quinbase about two years ago from Replit. I was a go to market lead at Replit and really helped launch Replit Agent, which took advantage of tool calls, enabled developers to really manage packages, deploy apps in the cloud inside a background around kind of Agenetic app deployments, joined Coinbase, and was working with a couple other people on the team.

We were brainstorming around kind of how do we attack the AI vertical and what problems can crypto really uniquely solve in this space to really improve the user experience for language model based applications or agentic apps. And so we really set out to solve three main problems that I'm that I'd love to dig into. And I think we've done a pretty good job of doing that. So the first problem that we recognized was around tool usage for for language models.

So traditionally, right, if you want your agent to do anything, right, in the real world, it needs an API key. And so we think about the agent stack, there's three components. There's the language model, which does the reasoning. There's the context window, which is really the memory of the agent, and there's the tool calling layer. Tool calls let agents take actions in the cloud, retrieve data, and really just take actions in the real world.

But the problem is that there's a lot of friction to set tool calls up for agents, for people who are nontechnical or for just even for developers. Right? A human needs to go to an API key dashboard, put a credit card in. They need to drive an API key, go back to the language model and configure that tool call, and then repeat that process for every tool they wanna give the agent. They have to go through this whole onboarding kind of account creation process.

And, you know, a great agent might have access to, like, 700 plus tools, and they need to drive 700 plus API keys. So, really, our big audacious goal was to kill the API key, replace it with a signature that happens at the client side, the wallet, and then the signature authorizes payments to let agents basically discover, reason about, and pay for tools at request time. It was really the first main problem that we set out to solve was was abstracting all of the human loop friction around tool usage for agents.

The second main problem is around economic reasoning. So if you give an agent an API key, it knows it can call that API and so often kind of spam it and retry it, but it has no idea how much that API costs. With x four zero two, it gets this four zero two status code response that gives it pricing information and context. So now it can think economically about the tool calls that that it's making. Right? And this is important because money is a really great alignment signal for kind of agentic workflows. Right?

Money sounds like a dirty word in Silicon Valley or or broadly, but really money is the coordinating principle that brings kind of the global economy together and aligns the global economy. And as such, it's a great alignment scaler for agents or great reward function for agents as well. Often, say, an agent that's aligned with an a user generally is an agent that's aligned with the user economically.

So giving agents real time economic data, giving them the ability to reason about marginal costs, qualitative trade offs, and other economic features is another key aspect of the x four two protocol.

**Zach Fowler** (5:02)

Did we get the third one?

**Kevin Leffew** (5:04)

Yeah. Third one. I didn't know you wanted to follow-up. Was like, go ahead. I can run through it. So the third piece, also really interesting, is enabling autonomous run times. So we know an agent can spend with an API key that's attached to your credit card. Right? That's easy to configure, but it's really difficult to enable an agent to earn. To basically enable an agent that earns, it would need to have access to a bank account. The bank account would need to be spun up by a human. Right?

It'd be attached to a human's credit score, so we need to a c ACH transfer or wire transfer money into that bank account. Then the agent will need to spin up a virtual browser, pretend to be the human, authenticate into that bank account. And then when kind of reasoning economically, right, it it'd have to know how much it has in the bank account and then reconcile how much it's spent on credit cards. Everything that's spent on credit cards is built in arrear.

And so they don't have to basically forecast, okay, how much I think I'm gonna spend on credit. And really kind of all of this complexity is really just context split for the agent. Right? And the wallet as a cryptographic primitive is a much cleaner construct that enables agents to both earn and spend in a single runtime and then reason economically in real time without all this, like, dislocation around, you know, billing in arrears and credit and having to ACH transfer money to the account.

With a wallet, they can they can spend and earn real easily. Mhmm.

**Drew Rogers** (6:20)

Mhmm. Kevin, I wanna bring this okay. That's solid foundation. We're solving these problems. It feels like in some way, you've you've solved most of them, maybe to various degrees. If you scroll through your x feed right now, it looks like like you're boots on the ground talking with different teams. It's funny you're how do you say the name? Orthogonal? Do you know how to say that? The AutoY Combinator. I yeah. I saw them I saw them separately.

Where where do you think about this now from, like, a go to market perspective working with these teams, working with the greater, like, AI community to kind of facilitate this this transition into the next version of the Internet?

**Kevin Leffew** (7:00)

Yeah. That's a great question. And look, I think this kind of hits on the point that's so important is that x four zero two needs to be open source. It needs to be open governance, and it needs really broad collaboration amongst different players in the tech community, which is why we formed and moved x four zero two to the x four zero two foundation under Linux.

This aligns it really closely with the MCP standard, and it also, you know, drives kind of a more trusted open source governance model that we think is needed to really light up the four zero two status code. And so we brought on partners like Cloudflare, AWS, Google, Stripe, Visa, Mastercard, lots of great names in the x four two foundation. If anyone on the call is interested in this space, the you can join. You can go to x42.org. We've opened calls, big open community. There's a Slack channel.

And it's a really exciting space to build. What's the what's the benefit from this being open versus closed or, like, making this move? What was the strategy there, and why is it better in your view? Yeah. So this really enables open agentic commerce. So, you know, open source infrastructure is critical. Blockchain itself is open source. But when I think of x four zero two, what the analogy that I often make is Kubernetes. Right? So Kubernetes standardized how containers orchestrate in the cloud.

It was originally kind of incubated or written by Google, moved to the Linux Foundation, and then in the twenty ten decade was adopted by all the major cloud providers, AWS, Google, Azure, Heroku, DigitalOcean. And what it did is it massively expanded the total addressable market for cloud compute in the twenty ten decade.

I think there's something analogous that's happening here with AgenTic payments in x four zero two, and we really want x four zero two to be that open standard with the trusted open foundation that any company, any organization can build on to facilitate agent native payments over HTTP on the web.

**Zach Fowler** (8:50)

Okay. So bring us in a little bit to the current moment. So we've we've kind of hit these three points. Why open source, OpenAgenta Commerce? You know, when we when we spoke initially, we were even internally thinking about some of the use cases that, as a media company, we were looking at with x four zero two and still are for, like, Internet payment usage for content and insights and all these different things, which we'd still love to talk to you about.

But bring bring us a little bit into, like you you seem to be, like, the first line of defense on the go to market team of, like, hey, we have this really cool business or use case or brand, like, they they want to talk to someone, and, you know, they're they're talking to you or your team. Like, what what are you hearing right now? What are what are the common, you know, eyebrow raising use cases or or types of brands that you are really seeing this type of infrastructure become very, very powerful for right now?

Because I think that it's it's easy to think three, five, ten years from now and start getting really excited. But Yeah. Specifically, like, you know, q four going into the end of the year, like, what is happening currently that is driving volume, is getting is getting, like, real utility?

**Kevin Leffew** (10:08)

Yeah. So I think when we think about what's actually useful today, these are all kind of, like, AI native APIs or tools that either let a language model take an action in the cloud or retrieve data for social feed for creative use cases for trading. And so, you know, inference tends to be actually the dominant use case category for x four zero two today, and that makes sense. Right? The inference is really kind of the lifeblood of the agent. Right?

The inference is all of the thinking, all of the output, all of the reasoning, and then we see a lot of usage around tool calls. But it would make sense that inference is kind of the the highest adopted category. We've brought on some really great AI native API customers. What I think is awesome is that this shows that x four zero two is solving a real problem for AI native companies. Right? They're adopting stablecoin treasuries.

They're adopting and accepting stablecoins for payments because it's enabling kind of open tool discovery. And these are companies like Browserbase, like Exa, like Browseruse, like Appify, like massive.com, which which we talked about that I that I just onboarded. But all of these companies are now basically supporting x four zero two within their API construct, accepting stablecoins, and they're using that to solve friction points around agentic tool calling.

**Zach Fowler** (11:18)

I kind of see this. I know the checkout and how we think of the checkout is really shifting, but it seems like this x four zero two seems like one like, can insert this into existing systems just like you might insert a payment method into a checkout, so you're not having to, like, completely rip and replace everything. So if a company's already thinking agent natively, you can add this to the stack without really having them rip and replace.

Is that is that kind of a paradigm that makes sense in how you're positioning this? I really liked the I'll pull this back up, the Coinbase payments image that I came across here where you're kind of seeing the infrastructure, where does it meet stablecoin settlement. I don't know if this is a helpful, like,

**Kevin Leffew** (12:02)

visual as we talk through this. Yeah. It's a great visual on the Coinbase side. But, yeah, the the goal with x four zero two is to really make payments simple because what we're doing is we're pagating any call over HTTP with the four zero two status code. So I'm sure we didn't really get the chance to go into this, but the 402 status code was built in the web, I think, in 1997 as part of the HTTP, I think, 4.1 spec.

And what it lets you do is really kind of restrict access to any endpoint on the web by requiring a payment. Stablecoins are a great way, and also credit cards are a great way to enable access to these resources. I think traditionally humans haven't really thought much about micro payments because there's, like, a lot of kind of cognitive overload and having to think about these small, like, 30¢ transactions.

But agents are actually rational economic actors that can ingest lots of information, lots of context, and execute micropayments at scale. And so I think we're really unlocking this kind of new category of commerce around tool calls and kind of API native micropayments, which is quite exciting.

**Zach Fowler** (13:06)

The the use case for micropayments, cognitive overload, like, what are some of these examples that machines and agents really love with micropayments that even thinking about them as human beings, like, it's it's I think sometimes we try to apply our human purchasing behaviors or economic activity to these types of conversations, it doesn't really work. Bring us into the types of use cases or behaviors that really drive this being special for machines and and streaming

**Kevin Leffew** (13:41)

money and and, like, what does that what does that look like? If this is a good if this is a good transition into what you have to share, feel free to pull up your screen, we'll we'll get it up. Yeah. Definitely. I can I'll I'll share my screen in a second. But to answer the question, basically, you know, we think about machine economy. It's API driven.

So this tends to be things like, you know, an agent that needs to set up a web browser or do some kind of, like, specialized deep web search or retrieve data via an API or maybe generate a video, you know, pull data off x. All of these actions are important because when we think of a language model, right, it's trained up to checkpoint. It's trained up to a certain point in time, and it then it doesn't have access to a real time context and real time data.

So any any way we can bring kind of deterministic real time data into the language model context window, that's that's when these use cases become valuable. Happy to share share my screens here and kinda walk through a quick demo. So I'd request this one on x earlier. What I have here is I have a Claude code set up with Coinbase payments MCP. So, really, it's it's pretty easy to attach a wallet into Claude. Right? Halsey, I think talked about Paybox previously. There are tools like agent cache.

There are lot of different ways to bring a wallet into your kind of native AI environment, whether that's Codex, Claude, Hermes, OpenClaw, whatever. So I can say, like, you know, can you show my wallet via payments MCP, which is one of the Coinbase tools that we have? So it's gonna pull that tool. We're allowed to spin it up. And then we can see here I have, like, you know, $90. I've set a bunch of micro payments. I can set all of these limits. So I can say, only want this agent to spend $5 max.

We'll move this up to 10, and we'll say, like, max $5 per API call. So we can set spending limits on the client side, then it'll require approval inside of Claude. But just to give kind of a quick demo here, I have this printed out. We'll go ahead and run this. We'll say you're an autonomous equity research agent with a wallet available through payments MCP.

Your goal is to identify US stock, making unusual, interesting move today, investigate what's driving it, determine whether the move appears justified by the available evidence, and then it's gonna pull real time context data, in this case, from massive.com, which will give a kind of historic price data, RSI data, MAC data, all kind of the signals that you'd want as a trader. And then we use Apify to pull social data around that as well to really kind of validate

**Kevin Leffew** (16:17)

just to kind of begin searching and find the market signal. And so what I wanted to showcase here is this is really the foundation for what we might consider to be kind of an autonomous trading flow. I think this is a super interesting area to start developing in now.

I think, you know, when I think about what's happening kind of in the financial sector, right, if if language models can disprove the Jacobian conjecture and solve the Ryman hypothesis, odds are they're gonna be pretty good at enabling trading workflows for the average kind of retail user. And so I think, you know, figuring out ways that we can kind of prompt optimize trading strategies by giving agents the right tools at request time and the right data, It's gonna be a really fun and interesting place to build. K.

So

**Drew Rogers** (17:00)

it has And then quick question here on on my part. What would happen if you didn't tell it to go to massive or if you didn't prompt to go to ape Apify?

**Zach Fowler** (17:10)

Do you think that is that provides maybe a question earlier with discovery a little bit? Like, what's where does the discovery layer hit?

**Kevin Leffew** (17:17)

Yeah. So there's a couple ways that it could discover x four zero two resources. It could use the bizarre, which will kind of just spit out a list based on a semantic search around different tools that might make sense. But I've also seen agents use their own internal web search API, go to a company's doc, find the x four two endpoint, and then access it that way for endpoints that aren't even available in the bazaar.

So I think there's a lot of kind of conversation around how agents make decisions around tool calls. Again, I think it's super interesting to kind of explore how that happens. We can run eval tests to determine, like, what kind of tool calls agents like. But with x four zero two, which is bizarre, it's getting all the pricing information so it can really reason kind of economically from a marginal cost perspective, which is quite interesting.

This gives us I think about it from the perspective of, like I've always kinda struggled with the idea of, like, SEO. It's like, oh, the answer that you're getting is whoever just spent the most time, like, putting together resources.

**Drew Rogers** (18:10)

And it's, like, kinda gaming the system to some degree. It's, do we want that when it comes to, like, an agentic engine or the resources that agents find? Like, is it finding the best possible tool, or is it just finding the most popular, most optimized tool? With the biggest pockets. That's like yeah. Though whoever has the biggest pockets to go out and rank high in the bazaar. I mean, maybe it doesn't work like that, but I think that's that's some of the attention that I I I feel.

**Kevin Leffew** (18:37)

Yeah. I agree. And so really the goal of x four zero two is we wanna drive a lot of experimentation around what we call, you know, tool reputation or agentic discovery. We think that this, like, isn't an area that's, like, should be solved at the protocol layer just like search isn't solved by the HTTP spec. Right? X four g just extends HTTP. So we want things like search, agent discovery, reasoning to happen at higher levels of the stack, But it's a really interesting place to build. Alright.

So what tell us again like, finish the demo here. So what what did it find ultimately? Yeah. So it's pulling it up here. We can kind of see the different tools that it paid, how much it paid for it. I asked it to basically give Rational for why it paid for this tool and what it learned. So it pulled a snapshot of the top stock gainers today from massive.com. It scanned the whole market for an unusual move. It found ChargePoint, this case doing, I guess, is 75 percent increase.

You know, then it found the immediate news catalyst for what drove this. So we kind of see here different tools that it called, eventually started running Apify to different to fetch different datasets kind of around this product as well. But we chain multiple tool calls together, and, basically, you know, the it discovered, okay. ChargePoint's moving. Why is it moving? What's the evidence supporting it? What's complicating it? How did it actually doing social research kind of change the language models hypothesis?

What's their confidence in this? And then it prints out a summary. And so, you know, what what I'm not doing here, which we could do, is actually tie this into an agentic trading workflow. So And you can imagine now if you can pull all this equity information at request time with x four zero two, it creates a pretty good real time context setup for actually executing trades and trading strategies, which is quite interesting.

You know, I could also ask it to say, you know, create a video in news style report using stable video. Right? So I can also take this content and then now, like, make a meme around it or something. So there are all kinds of, like, really interesting trading creative use cases that are made possible when we can chain all of these different API requests at at basically request time.

Or otherwise, right, for me to do this, I would have had to go get, like, four different API keys, put my credit card into four different places. But with x four zero two with the wallet as a client construct, I get access to 7,000 plus tools that the agent can reason about and leverage at request time.

**Zach Fowler** (21:04)

So interesting. Question for you. Like, there are so many different pieces to equipping this agent flow. There's so many different pieces of this transaction, like information being supplied, insight being supplied, news sources.

Like, I mean, even in this one single demo, like dozens of these that are being supplied, like, when you think about the landscape of information, media, news, data, whatever, like the source material, where do you ultimately see the most value coming from, at least in the context around trading, financial transactions, whatever the case might be. If I'm wanting to place a trade or look up like market related information? Like, where do where do you see the most valuable context coming from?

**Kevin Leffew** (22:00)

Yeah. It's a great question. Look, you know, not not financial advice. There are a lot of different sources that I'd look at. So I'd look at technical indicators like RSI and MACD, and then I'd also look at social indicators. I try to combine those together to let the agent model reason about a trading thesis based on technical indicators, news, and social feedback. That's been my strategy to date. But, you know, I'm interested in So social means like sentiment? Like public sentiment?

Like, what are people tweeting about, you know, this move? You know, what are they posting about it? I think that that agent that kind of context is organic and can be really rich for kind of the agent, reasoning as well. Interesting. Be careful. Be careful. You know, a lot of this is experimental and new. Cool. So

**Drew Rogers** (22:44)

No. Where my head goes, though, is okay. So we're we're looking at market sentiment. How do we know if the sentiment's coming from humans or coming from bot activity? You know, I think about somebody that if you're wanting to push a certain agenda, could you spin up 10,000, you know, x agents that are going through and, oh my gosh. The sentiment around this token is so you we've know, kind of already seen that already in in crypto to some degree.

And I think this is something that we even just think about is, like, what role does what humans actually think and what humans actually said have in in the machine economy? How do you verify that information? How do you know that it's true? How do you know that Kevin or that, Brian Armstrong said something versus

**Kevin Leffew** (23:26)

bots on Twitter talking about, you know, coin or something? Yeah. And that's where I think kind of selecting the right APIs. Like, you know, you probably would wanna underweight social sentiment, and you'd probably wanna overrate overrate, like, I don't know, you know, PR Newswire and things like that, like official news sources. And so you'd wanna feed all that context to your agent to get it thinking.

I think what's also important, right, is, like, you you kind of wanna be the human in the loop to verify all these trades initially. But where I see this going is, like, autonomous execution over time. And I think that there's gonna be a lot of like risk taking and kind of interesting activity that where the we're really just kind of like handing the keys over to the agent. And so I think we're the early innings of that. Right? It's like, you know, you wanna really limit your exposure.

So like with payments m c p, I can give my wallet like $10, and the max it can spend is $10. So I can kind of limit like collateral damage, where like if I give it a credit card, maybe it would like not spin over a trip because we crack up a huge credit debt. That's what I like about Oh, I have a $10,000 credit limit. Yeah. I mean, we could. So that's what I like about, like, kind of the hard limits of cryptography.

Like, I can really be sure I can be sure on the limits that I set, then, like, there's this max, like, okay, it only has $10 in it. 's supposed you you know, lose or a $100 or whatever. But yeah. I think it's a really interesting space to experiment in.

**Drew Rogers** (24:44)

Last question from my side. Oh, go ahead. Go. Go. I'll wait for mine. I was gonna say we we we had a question come in prior to the screen stream, and I just had it up. Let's see how fast I can get to it. Oh, yeah. Here it is. So somebody had asked with Coinbase for agents connect connectors, if the human closes the computer session, does the agent keep the ability to place the orders, or does the trade authority you know, does it kind of persist through when that window dies?

I think it kind of ties into what you were kind of just talking about too with, like, where's the human in the loop? Is it the initial approval? Does it, you know, retain that authority to work autonomously? Is it a little bit of both?

**Kevin Leffew** (25:28)

Yeah. And it depends on the agentic setup. Right? So that question's for Coinbase for agents, which is giving kind of API access to a Coinbase account. You know, you could also have an agent trade on chain. You could have an agent do tool calls over x four zero two. But then when we think about what makes an autonomous agent, right, that's gonna be something where, you know, you're dangerously skipping permissions, and you're running it in some kind of, like, hosted environment.

And this is probably more of like a Hermes or an OpenClaw agent versus something like CloudCoder or Codex. But, you know, what's what's what's happening here, it's it's really how the client configures the agent. And so a lot of this is configurable from the client side. If you're running this in, a long running, you know, OpenClaw Mac mini that doesn't turn off, well, then it can keep running, keep pulling more con more contacts.

You know, generally, I like to set permissions and, like, have to approve transactions. But, you know, as a user, I can actually delegate that to the agent. I can basically give it, you know, my keys.

**Drew Rogers** (26:23)

Do

**Zach Fowler** (26:27)

I get to go now?

**Drew Rogers** (26:29)

My turn. Yeah. Go ahead. So

**Zach Fowler** (26:32)

we talked about Linux Foundation, X4R two kind of living in this, like, open standard. What do you, as a commercially minded participant within Coinbase like, what is the commercial advantage or the commercial centers where Coinbase really has an amazing advantage as a I don't even know what the word would be, like contributor or like influencer around x four zero two. Where are the profit centers for the business? And where do you see this, like, being a really unfair advantage to Coinbase as a as a company?

Like, what products or, like Yeah. Profit centers

**Kevin Leffew** (27:10)

do you see? So we we don't have any advantage on x four zero two. I think the analogy is kind of like MCP for Intropic. Like, we developed x four zero two. We realized it's a really good way to do paid tool calling, and we wanted to release that to the world. Where we do, I think, have an advantage is on trading. And so I think tool calling at request time is really important for trading agents and making trading agents accessible to general users. Right?

Like, you know, to like, I don't think for example, my sister is gonna be go going to 700 plus API key dashboards, getting API keys, configuring them. Right? But, you know, maybe she would use Apple Pay to on ramp a $100 in USDC and do paid tool calls. So I think what this enables is more user friendly experiences for agentic trading workflows. I think trading is where Coinbase has a a structural advantage.

**Zach Fowler** (27:57)

And, obviously, that's a big, you know, revenue center for us. So Yeah. So it's it's widening and and growing the pie of trading, which will inherently help Coinbase in in the core business.

**Kevin Leffew** (28:07)

That's right. Yeah. And I'm interested in, like, what tools are valuable for trading agents and bringing those into our ecosystem and making those accessible specifically on base in our case. So, like, how can we create, like, a base state of ecosystem around tools that give agents the right context,

**Drew Rogers** (28:23)

the right abilities to take actions in the cloud, retrieve data, and execute effectively as as trading agents. That's so cool. So cool. And I got I got two more questions for you if if you don't have a hard stop. One is Yeah. What did that did it make a video? I'm, like, sitting here with a open open one back up. I'm like, I I can't end end the stream right now. And I because I saw it say, oh, I don't have this tool, and then so I'm curious what you chose. Yeah.

Second was we would ask Halsey about this, and then I saw you tweet about it as well with fall and generated video live streamings. Any thoughts to unpack there? Like, why does that make you so excited? Like, what would you want to see with, like, generative AI live streams?

**Kevin Leffew** (29:06)

Yeah. Yeah. So I'll answer the fall question first. I think it's so interesting. So fall released h three max director, I believe what it's called. And so basically, it can generate video faster than it can be streamed. Right? This enables these kind of, like, infinite live streams. So that is super interesting. Right? We saw, like, infinite cable or whatever it's called from Rick and Morty. Right?

There's been, like, similar experience in experiments in the past with, like, infinite Seinfeld and others, but the quality is getting better. And where where I think things get interesting is, like, how can we actually incorporate, like, payments into guiding the adventure around InFit livestreams? Right? So, you know, if we think about, like, you know, how human livestreams work today, like, you might tip a creator and be like, you know, do a backflip or, you know, do something cool. Right?

But we can we could imagine something analogous to, like, a highly viewed, you know, procedurally generated livestream where some human is, like, paying a tip. The tip is then waiting the context window, and it's, like, steering the generation of the video. And so, you know, I think that that's, like, a really cool concept that is now just possible over the last few weeks. So if I was a builder, you know, I'd be thinking about, like, that intersection between incentives, procedural video generation,

**Drew Rogers** (30:20)

live streams, and, like, how to bring stablecoins, x 14 crypto into the mix. Not gonna lie. I have a chat going right now that's, like, ideating, and I'm fighting a shiny Yeah. We're gonna

**Zach Fowler** (30:32)

we're gonna let Kevin get on with his day before we we take too much more. Yeah. Do you wanna Kevin, do you wanna show the end result, or did do you wanna pause there on the demo? So I didn't approve it. I just approved it now. I can share it later.

**Kevin Leffew** (30:48)

Okay. But, basically, what I what I prompted to do is create a new style, you know, taking all this data around this equity, like, you know, create a creative asset around it. And so I just wanted to showcase, you know, the advantage of, like, pulling data, generating media, or executing trades, like, all the different things that are possible with the request time tool request time tool calling with x four zero two.

**Drew Rogers** (31:08)

Well, if you see a if you see a clip that goes out, pay attention either all the way through Monday probably. There might be a clip that goes out at some some time that the ending of it is like, and we talked about the clip. Maybe that's a perfect, quote tweet of like, hey. And here's actually what to what it made if you're curious. So quote tweet it, put it in the comments, and, we'll help you out too.

**Kevin Leffew** (31:28)

Will do. Well, thank you so much for having me on. I hope you guys have a great Labor Day coming up, and excited to hop on and talk about the x four two ecosystem, AgenTic payments, AgenTic trading anytime.

**Drew Rogers** (31:40)

Cool, man. Awesome. Thanks for coming on, Kevin. Good to meet you. Good to meet you.

**Kevin Leffew** (31:44)

Cheers. Cheers.