Why AI Agents Will Need Their Own Bank Accounts — Lincoln Murr, Coinbase

AI Engineer · 16 min · 204 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:12Thank you all for attending.
00:14My name is Lincoln.
00:15I'm the head of AI product at Coinbase.
00:17And today we'll be talking a little bit about how we're building the agentic financial account.
00:22And I know what you're probably thinking.
00:24Coinbase, crypto, AI, a lot of buzzwords.
00:27But my hope is by the end of this conversation, you have an understanding of what agentic finance is and that it's a very real thing that's happening.
00:34You understand a little bit about what we've learned through this process of building this over the past 2 years and hopefully can take that and and distill some insights from it that are helpful for you.
00:44And I'm going to talk about a little bit about where I think we're heading with this whole narrative and how agentic finance will evolve over time.
00:50Cool.
00:51So, to understand first uh where where I'm going with this, I think it's important to look at what is agentic finance.
00:58And in my mind, agentic finance is anything that has to do with an agent that can manage money or has access to a financial account.
01:05Kind of obvious, but what does that actually mean?
01:07There's really two use cases that we're seeing.
01:10It's agents that need to pay with money and agents that need to grow money.
01:15So, the grow is a little bit more of a nascent category.
01:17So, I'll start by diving into the payment space.
01:20And to do that, I think it makes sense to start around 2 years ago with our first exploration into agentic finance, which was this tool we created called Agent Kit.
01:29The whole idea was prior to MCPs, really any sort of standardized AI tooling, we uh used the LangChain framework and added a wallet to that.
01:38And we gave AI access to money and crypto particularly.
01:41And at the time, this was a pretty novel thing.
01:43And I remember it was like 2:00 a.m. I was typing into my agent like, "Hey, can you send money somewhere?"
01:47And it did it.
01:47And I was like, "Whoa."
01:49These days that's not impressive at all.
01:50But at the time it was pretty fun.
01:52But as we gave this to our developers, we noticed that there is one core problem.
01:57We gave an AI access to money, but we didn't really give it anything to pay for.
02:02This was a bit of an error on our part, and it's something that we wanted to solve, and we were able to do it by standing on the shoulders of giants.
02:09How many of you are familiar with HTTP status codes?
02:12Hopefully, yeah, most of the room.
02:14Okay.
02:15So, for those of you who aren't, it's basically a standard that defines how websites communicate and how information is transmitted across the internet.
02:23One very popular one is called 404.
02:25That means there's an error.
02:26200 means that there's success, but there's one that actually has been in the internet spec for decades that nobody really ever implemented.
02:35And in fact, it was called the original sin of the internet by Marc Andreessen that we never properly implemented this.
02:43For those of you who know a little bit about what we work on, you would know that this is the HTTP status code 402, which means payment required.
02:50And the official definition of this in the internet spec is reserved for future use.
02:55Pretty generic.
02:57But we have a great opportunity here now with the inflection points of AI agents needing a clean, fast, efficient way to transfer money and engage in machine-to-machine commerce, and
03:06stable coins as a cheap, fast, efficient payment rail for us to build X402.
03:11It's our standard for internet native payments.
03:13This says it's new.
03:14I guess it's still relatively new, but it came out in around last May.
03:18And since then, we've seen some very strong adoption, honestly much stronger than I ever thought we would.
03:22When it came out in May, our goal was for by the end of the year to have 25,000 transactions made by agents.
03:27At the end of the year, we were actually seeing that every 17 seconds.
03:31And these days, we're seeing pretty incredible numbers across the board with tens of millions of volume and hundreds of millions of transactions occurring every 30 days.
03:39And if you'll notice from this list, you'll see that it's more than just crypto companies.
03:42You see Audius, Amex, PayPal, Visa, MasterCard, Stripe, all these big names in payments who are engaging with X402 because we're building it as an open standard where anyone can use this how they want to create a way for machine-to-machine payments to happen.
04:01So, I know there's been a lot of traction and comments around agentic commerce recently, especially with news and Grokbot and Stripe Link and this whole Amazon blocking agents controversy.
04:11And I think that's all very straightforward use case.
04:14And I can talk about agentic payments all day.
04:16But there's one that I want to touch on that I think is particularly relevant to microtransactions, which is something I'm very excited about with this whole agentic payments narrative.
04:24So, if you all haven't noticed, the way that we've interacted across the internet has changed over the past couple of years.
04:30Now, we're primarily doing it through chat experiences or Google's AI overview instead of going to the actual websites themselves.
04:37And while this is all fine and dandy for the consumer, it's resulted in a bit of a problem for these publishers who have previously relied on ads revenue and human eyeballs to generate money for publishing their content.
04:48It feels a little bit like we're in the limewire era of the internet where you can scrape all of this content for completely free and there's nothing really to be done about it.
04:55And recently we saw that bot traffic is now eclipsed human traffic on the internet for the first time and this number is only going to increase as we see more people using AI and these traditional
05:06web services and and websites be disintermediated.
05:09So, how do we solve this problem?
05:11There's a very compelling opportunity here that's just starting to play out and it's it's really coming to life over the past couple of months for websites to gate their content behind a very small microtransaction
05:22specifically for agents.
05:24So, the idea here would be if I publish a Substack and I don't want an agent reading it for completely free, I can charge them 1/10,000th of a cent.
05:32It's pretty trivial to the user.
05:34It should ultimately be fully abstracted away in the same way that your AI subscription handled this for you, but it's a meaningful revenue driver for these companies.
05:41And we're going to start seeing a shift from this attention-based ads economy online to one that's primarily oriented around micro transactions and and utility.
05:50And if you don't believe me, I I encourage you to look at AWS and Cloudflare.
05:54Both of them, they're these massive content delivery networks.
05:57Together, they control somewhere in the neighborhood of 40% of all internet traffic, and they've both recently enabled X402 micro transactions to gate any website that their publisher is creating.
06:09And though this hasn't seen very strong adoption just yet, it feels inevitable to me that over the coming months and years, this will be the new way that we have companies
06:17differentiate and continue to undisintermediate themselves in some sense.
06:23So, payments.
06:24Make sense.
06:25We're seeing some pretty strong traction here.
06:27All is going well.
06:28But what about this whole grow with money thing?
06:30What could this possibly mean?
06:31Well, it's more agentic trading and investing.
06:34And the first time I heard about this, I thought it was the most ridiculous thing in the world.
06:37Why would I ever give my ChatGPT access to my bank account?
06:41Or why would I ever tell Grok to make me money and actually expect it to work?
06:45But after digging into it a little bit more, we found really three core use cases that exist today.
06:50I'm sure there'll be more going forward.
06:52And the the core differentiation between this whole Grok make me money thing and what's actually happening is the human just providing a teensy bit of agency.
07:01And what I mean by that is the the human still sort of directing the agent to automate these tasks that would otherwise be very difficult.
07:08So, say for example, I just want to have my low-risk, you know, 80% S&P, 5% Bitcoin and, you know, 15% T-bill account fully managed and handled for me by an agent.
07:19I can just tell it to do that, give it authorization with all the appropriate guardrails and whatnot, and let it run.
07:24And I don't have to think about my money.
07:26Similarly, we've seen autonomous execution be a pretty compelling use case.
07:30I'm going to go into an example about what this means later on, but you can imagine if you're interested in trading the the straight-from-the-news and what's happening with oil.
07:37You could either be sitting in front of your keyboard at 3:00 a.m. watching these charts and watching the news, or you could tell the agent, "Hey, this is what I think is going to happen in these different scenarios.
07:46If this happens, just make that trade."
07:49And the third one that's been particularly compelling is the idea of taking a high-level natural language idea and translating that into actual financial primitives in in a strategy.
07:58Invest in space.
07:59Invest in quantum.
08:01Give me a low-risk portfolio.
08:02Invest like Warren Buffett.
08:03All of these things for people who have these ideas, but they don't necessarily have the time to execute on them, nor do they exactly know how, is a very compelling opportunity for agents to come into the foray.
08:14So, I'll give you one coherent example of something I was actually doing a couple weekends ago.
08:18So, you you all may may remember that Daria put out this post, "We must paste the frontier."
08:23This was over the weekend.
08:24This was about a month ago.
08:26And I saw on X that Elon said, "Dario is right."
08:29I was like, "Holy That's the first time that these two have agreed on anything.
08:32That's probably going to be market-impacting news come Monday."
08:35So, I went into my agent and I said, "I think that blank stock" You might be able to guess what it is, but I'm just not going to not going to go there.
08:43I think that blank stock is going to go down.
08:44And because we're now entering this era of 24/7 markets, I was able to put a short on these positions over the weekend just by talking into my phone.
08:52Could I have done this by sitting in front of a computer and then pressing all the buttons manually?
08:55Yeah, totally.
08:56But having it completely handled for me was so easy.
08:59Let me go about my day and not really think about it.
09:01And since then, since this worked fairly well, I now have my agent searching over the weekend for any news that might be market-breaking or blockbusting or whatever it might be, and it lets me know anytime that happens such that I can make a trade off that going forward.
09:14And is this specific opportunity going to exist forever?
09:17Likely not.
09:18But the whole idea here is that it's so simple and so straightforward for me to just talk my ideas to life that there's a very compelling use case for for agents around this whole thing.
09:30This is another way to look at it.
09:31This is kind of how we think about building this AI-enabled financial account, and And really, you know, breaks down into advanced traders, and also more like retail consumers, and and what exactly we want to do to provide them with the best-in-class experience across
09:44uh across the spectrum of capabilities that we can offer.
09:48So, all that being said, we've we've been building this stuff for about 2 years now, and we've had a lot of takeaways and lessons that have come into play,
09:55and I think are pretty broadly applicable to building for agents and building for this new agentic era.
10:02The first is that trust and autonomy are earned one set at a time.
10:05The most question we get when we tell people that we're working on agentic finance is, "What if the AI goes rogue?
10:11Why would I give money to an AI agent?
10:13How can I trust that this thing is going to work?"
10:15And it feels very analogous to me to the way that we've seen self-driving cars start to proliferate.
10:20A couple of years ago, you'd be insane to try to use lane assist to parallel park, but now you go to San Francisco, you jump in a Waymo, and it feels completely natural.
10:28And we're seeing the same level of progression with the way that people view AI having access to money and finances.
10:36It starts with complete hesitation and and no real thought whatsoever to the concept.
10:41Then, people come around to the idea that, "Well, I mean, my AI is already chugging through credits.
10:45That's kind of a form of money.
10:47What's the big deal if I gave it a couple dollars to pay for some microtransactions?"
10:51From there, you'll notice that these models are quite conservative and and aligned to the the users' needs, and they don't really do anything crazy.
10:59Anecdotally, over the past couple of years, we've never seen any situation where a model goes and and blows up a user's entire account in a way that's completely uh un- unfixable.
11:09So, then we get to the point of a user feeling comfortable with the right guardrails and rules and safety policies and procedures in place to give their agent access to an isolated portfolio that they can attempt to grow however way they see fit.
11:21One way is trading, but another is by offering a service and seeing if that agent can go make money by offering whatever it's doing and getting paid for that via these X402 rails that we're talking about.
11:31You can see how from there it just continues to to go on and on.
11:35The other thing that it has been particularly interesting to see is that especially in financial services, AI and agents are not the product.
11:42It's actually the enabler.
11:43And this is a screenshot from one of our internal AI products called Coinbase Advisor.
11:48It's actually an SEC registered AI AI financial advisor.
11:51It's the the first in the world of its kind.
11:53And what we noticed is that AI and this concept of an AI giving financial advice, even though we've had a team of asset managers who are tuning this model,
12:02that's not the value prop.
12:03The value prop here is that we can scale this advice to hundreds of millions of users, provide it for completely free, and give them the ability to talk to a trusted regulated entity 24/7.
12:12People actually don't care, don't even necessarily know or want to know that there's AI going on behind the scenes.
12:17And though it may be obvious with the chat button and whatnot, I think this is going to lead into more generative UI experiences that feel pretty traditional, where you have agents behind the scenes and whatever sort of interface you want on the on the front.
12:29Another thing we saw too is this whole AI chat experience.
12:33Though it's great for these generic use cases, it's very obvious that there's this blank canvas cold start problem that happens, where users come in and they have absolutely no idea what they can do with this stuff.
12:44So, we started using these widgets as a way to show the user proactive insights and and opportunities that exist and and what they can start using these for.
12:53And we've seen really strong conversion through this surface specifically because of the general nascency and lack of education around agentic finance and what these opportunities are.
13:02And I suspect over time as people understand it, and you see this a lot too with the, you know, Google Gemini and OpenAI ads, they're all trying to subconsciously
13:09drill into users the variety of use cases that you can take with these products.
13:14Lastly, all these next-gen agents, particularly Grok Bot and Muse, they all have a micro VM.
13:20It's like a little personal computer that they can run things on, have storage, have access to a file system.
13:26That's been huge for our development process.
13:28We are now able to create CLIs and MCPs to let users connect their Coinbase account into these agents.
13:35Nothing special.
13:36What is special though is that we get requests all the time for things like, "I want to do paper trading.
13:41I don't want to give my agent any money first.
13:42I just want it to trade entirely on its own."
13:44Or "I want to backtest."
13:45Or do all these different things.
13:47And prior to these agents being a thing, we would have to go to our internal brokerage team.
13:52We'd have to build within this random repository.
13:54All these changes in in big tech things would have to persist and distill and percolate and and just a lot of work.
14:00But now we can just write a skill to write the paper trading code onto the user's machine.
14:05And while this probably isn't the most uh the most efficient way to do it, it allows us to execute and build and iterate at the pace that our customers are demanding
14:14without having to worry about all of the implementation details and all of the dependencies and traditional software cycles that might come into play.
14:21So, with that I want to leave you with a couple thoughts about where I think the space is going and why agentic finance is going to be such an interesting opportunity.
14:28My core thesis here is that financial independence leads to greater autonomy for these agents.
14:34And you can imagine that in the same way when a human has access to a card on the internet, they can do so much more than they can when they're just browsing and don't have access to pay for any things.
14:43I think the same exact thing is true for agents.
14:45And we're going to get to a point where we can see these agents interacting over long periods of time.
14:50And maybe they're just bringing in humans at the edges when they have a specific regulated task that needs to be take place.
14:56Like forming a company or getting a legal opinion or any one of these tasks that is is going to maintain its its human oriented factor for the foreseeable future.
15:05There's a really interesting uh thing that we've been thinking about as a thought experiment, which is what happens when an AI is able to sustain itself.
15:12You give it access to money.
15:13You give it access to a wallet and you allow it to start earning.
15:16Can it start to pay for its own inference?
15:17Can it start to pay for its own compute and its own upkeep?
15:20And does this start to feel a little bit like a like a proto-digital life form of sorts where this AI is now able to just exist and subsist based on the the the quality of the services that it's providing?
15:30And though I think this is a long way away, I think it's very possible to see how this concept comes back to life.
15:34And within the crypto space specifically, there's this concept of a DAO.
15:37It's a DAO, decentralized autonomous organization.
15:41And it's been an idea for for almost a decade now, but there's never been really any interesting implementation where there's automation at the center and humans at the edges.
15:48But I think now is the time where this is finally going to become a thing.
15:50You have an AI that follows a set of rules and has access to a treasury to manage at that core, and it can hire out humans to do other tasks at the edges, but it's really this self-sustaining
15:59uh organization that that uses AI and agentic finance to make that all come to life.
16:04Another example being self-driving car that owns and run itself.
16:07Give it access to money, let it take people on taxi rides, use that money to pay for its operation and maintenance and upkeep and whatnot.
16:14It It's a very interesting thing, and you know, who knows how long it'll be until that comes to life, but it feels inevitable at some point.
16:20And the the whole idea here being that agents are just going to be start transacting at the speed of machine-to-machine commerce.
16:25They are going to be both paying and earning and building, and ultimately I think it's going to be the primary way that we see money move across the internet in a velocity as which we've never really seen before.
16:35So, that's all I had.
16:37Really appreciate everyone's time today.
16:38If you're interested in in collaborating or jamming, I would highly recommend you play around with X402 and the Coinbase MCP and all of our fun products.
16:45I promise you they're much easier than you think to get started with them.
16:48And uh yeah.
16:49Thank you all.
16:50Thank you all for your time.