[BidClub_]
David Senra · · 60 min

Why Now Is The Golden Age of Building | Brian Armstrong, Coinbase

David SenraBrian Armstrong

CryptoAI & SoftwareFinanceCompany BuildingTechnical
YouTube ↗
TL;DR
  • Coinbase’s AI productivity gains have not yet materially lifted revenue. Armstrong favors a lagging-indicator explanation—perhaps another six months—over a competitive “red queen race” in which everyone improves without gaining ground. His explanation remains a guess, making revenue conversion the unresolved test of the productivity story.
  • Armstrong is reorganizing work around persistent agents, company memory, and human decisions. He runs seven or eight lead agents plus a chief of staff, with leads spawning roughly three to five workers when active. He describes their capabilities as “L five, L six,” while humans retain final decisions and code reviews.
  • Agentic finance is Coinbase’s most expansive—and speculative—growth thesis. Asked what share of transactions would be agent-to-agent in ten years, Armstrong answers “Five nines”—99.999%, “or whatever”—and thinks the tipping point could arrive sooner than expected. Credit cards provide backward compatibility; stablecoins supply his envisioned real-time payment rail.
  • Traditional finance remains the larger competitive target than rival fintechs. The conversation puts fintech at about 4% of global financial-services revenue against traditional finance’s 96%. Armstrong thinks agents could create a compelling reason to switch accounts: “Crypto are the rails. AI is the interface.”
  • The proposed consumer product moves from executing orders to generating and implementing financial ideas. Armstrong’s progression runs from buying Meta, to building a 10% allocation over time, to translating a thesis into an investment basket. The destination is “a private banker in their pocket,” including proactive ideas, tax savings, and better use of idle cash.
  • Armstrong sees clearer commercial value in financial infrastructure than generic AI coding tools. Coinbase Developer Platform has seven of the largest banks building on it, he says, although the business remains early; he envisions something “kinda like AWS for us.” Generalizing an internal AI tool could require 10 or 50 times the work, while licenses, funded accounts, liquidity, and network effects remain harder to reproduce.
  • The golden age of building still has operational and physical constraints. Armstrong wants fewer third-party software vendors to reduce security exposure, while AI agents review every production-bound line of code. Humanoid robots would unlock the next labor expansion, but New Limit is waiting: “It can’t do barista yet,” so laboratory deployment is premature.
Digest · the substance, structured for research

1. Deliberate self-talk turns dissatisfaction into action

  • For Armstrong, repeatedly writing a goal keeps it salient rather than manifesting it through the universe. In his twenties, he wrote, “I wanna start a billion-dollar tech company”; after moving to Silicon Valley at roughly twenty-nine, he reached that milestone within six or seven years. Updating neural-network weights is his metaphor for repetition, explicitly offered without claiming neuroscience accuracy.
  • Conflict aversion became a management problem because only the hardest unresolved decisions reached him. His repeated instruction was “I love conflict”—a way to reinterpret confrontation as meaningful progress. Where he once bottled up frustration, he now might tell a team its work misses the bar and needs changing by tomorrow; employees have sometimes thanked him for saying what everyone sensed.
  • Senra calls this David Ogilvy’s “divinely discontent.” Armstrong says he contacts thirty or forty people a day about problems in Coinbase’s app, yet describes his internal response as “Fuck, we gotta make this better,” rather than self-punishment. Even success produces “What’s next? What’s next?” Senra’s corresponding standard is to be “embarrassingly passionate” about the work.
  • Armstrong distinguishes temporary happiness from fulfillment earned through meaningful work, relationships, and progress. Coinbase’s mission of a fairer, freer, more efficient financial system is, in his assessment, only “one percent of the way there.” Difficulty belongs in the experience: an easy video game would be boring, and comfort alone would not leave him fulfilled.

2. Civilization supplies a model for building over decades

  • Armstrong keeps a printout of Civilization’s technology tree in his home office. Its progression from agriculture, writing, and pottery toward AI, biotech, and future technologies shaped his ambition to add “a leaf” and “expand it a little bit.” Its multiple victory conditions also matter: scientific or diplomatic progress offers alternatives to military conquest.
  • His concrete strategic lesson comes from fighting one civilization while another grows powerful on a different continent. Gifting technology to the distant power’s weaker opponent prolongs that struggle, preventing a future rival from becoming overwhelming while he finishes his immediate battle. The attraction is the game’s layers of resource allocation and interdependence.
  • Business scratches the same itch across decades rather than twenty-minute games. Armstrong initially preferred solo programming, then learned that greater ambition required recruiting and empowering teams—frustrating skills for an introvert who thinks doing it himself would be faster. Coinbase now has roughly 4,000 people. AI might enable a one-person billion-dollar company; he does not know whether that milestone has technically been achieved.

3. Persistent agents need both specialist models and institutional memory

  • Coinbase’s progression is from prompting for responses to assigning continuing responsibility. An onboarding agent can analyze country-level performance, identify five opportunities, propose starting with three, and seek approval. Pressed repeatedly by Senra on the quality of these suggestions, Armstrong places agents at “L five, L six” employee levels, rather than claiming they already run major divisions independently.
  • Expensive frontier models handle planning; subtasks can go to cheaper models, with an “auto” setting choosing for employees. Coinbase also fine-tunes smaller open-weight models on internal data. In some cases, a narrow model—for example, one drafting and filing compliance reports—outperforms frontier models while being faster and cheaper.
  • “Coinbase Brain” accumulates context for every employee, team, and GitHub repository. Inputs include A/B tests, user studies, incidents, meeting transcripts, code reviews, and policy documents. Nightly processing distills lessons into the relevant brains, which agents load into context before working. The aim is better alignment with company practices and more work completed correctly without repeated managerial correction.
  • Armstrong cautions that AI resembles humans as “a submarine is not exactly like a fish”: it works around the clock and can be copied, yet sometimes misses the big picture or feels derivative. Humans retain final decisions and code reviews. Their growing responsibility is cultivating the brains and resolving contradictions—including whether David’s guidance or Brian’s from fourteen months earlier should govern.

4. Decision queues become the interface for managing agents

  • Armstrong currently has seven or eight lead agents and a chief-of-staff agent; active leads may each launch three to five sub-agents. That quickly overwhelmed a chat-based workflow with context switching and walls of text. Coinbase is prototyping a prioritized inbox of decisions so five available minutes can unblock the most consequential work.
  • Product reviews can arrive as screenshots with a proposed next step, answered through approval, rejection, or brief voice feedback. Consistent button positions and keyboard shortcuts reduce overhead. A future layer might operate more like a VP: “I think we need to expand into Brazil,” followed by authorization and a check-in every three months. He presents that as a further capability, rather than today’s baseline.
  • His broader prototype maps priorities, projects, people, approvals, and critical dependencies. An envisioned AI might flag a weakly staffed project supporting fourteen downstream tasks, then recommend moving someone, replacing the leader, changing strategy, or narrowing scope. The manager’s decision would direct the resulting action.
  • One current lead handles retail-app quality. Armstrong mostly drafted an X post requesting complaints, then asked the agent to summarize the top five issues and launch workers against them. It assigned five. He believes a couple of fixes are almost live. Discovery is collaborative: sometimes he pushes on a stalled problem; sometimes an agent surfaces customer needs from the data.

5. Agent coordination creates a case for financial markets

  • Armstrong extends the coordination problem beyond one company. Individuals, teams, companies, and countries accomplish more through specialization and exchange; agents will need equivalent mechanisms. Markets and money let security, design, financial-analysis, and product specialists assemble into something more capable than isolated workers.
  • Senra pushes the scale implication: if agents can create copies and operate without a human at every step, why would their economy not dwarf one limited to eight billion people? Armstrong agrees. Asked separately what share of transactions would be agent-to-agent in ten years, he answers “Five nines”—99.999%, “or whatever”—while saying the tipping point could happen sooner than expected.
  • Credit cards initially matter for compatibility with existing purchases, such as shoes and flights. Stablecoins, in Armstrong’s view, become the real-time payment rail for the agent economy. His broader Coinbase formulation is to combine frontier technologies with better financial services, ultimately serving economic freedom: “Crypto are the rails. AI is the interface.”
  • Decentralization is part of that vision. Armstrong worries about one or two AI providers concentrating control or bias; small, customizable open-weight models and competing specialists preserve choice. He wants Coinbase to become the most AI-enabled financial account, with strict permissions and customers retaining control, applying the trust-building approach he says the company used for crypto.

6. Security vigilance coexists with skepticism about precautionary regulation

  • Armstrong says cyber incidents across crypto and traditional finance have roughly doubled year over year. Coinbase uses frontier AI agents to review every line of code before production and stores 99% of crypto entirely offline. Third-party vendors remain a concern because Coinbase has less control over them; rebuilding software internally could reduce both security exposure and spending.
  • Senra asks why entrepreneurs’ excitement contrasts with public hostility despite widespread AI use. Armstrong thinks cautious responses to change run deeper than poor messaging. His example is Amazon: people can worry about its effects on workers or local businesses while continuing to use it. Utility and concern coexist; adoption does not require unqualified enthusiasm.
  • Armstrong argues against regulating hypothetical harms before understanding the actual risks. He cites historical fears around bicycles, video games, and civilian cryptography, and points to nuclear regulation as a warning about suppressing an industry. When imagining an objection that AI could end the world, his response is skeptical: “There’s no historical precedent for that.” Senra agrees that humans often solve problems once confronted with them.

7. Internal AI tools face a demanding commercialization test

  • Senra contrasts Shopify keeping powerful internal tools with Ramp selling tools developed for its own needs. Armstrong says Coinbase has had the same debate. Making its prioritized agent interface useful across companies could require 10 or 50 times the work, including different security requirements and billing preferences. Open-sourcing it would still create maintenance and feature-request obligations.
  • Armstrong suspects open-source alternatives will limit how much value coding harnesses capture. Senra offers Bending Spoons as a model for applying proprietary tools inside acquired businesses rather than selling the tools themselves. Armstrong’s immediate preference is using specialization to improve existing products generating billions in revenue. He is trying to become more disciplined about launching additional product lines.
  • Financial infrastructure has already passed that test. Coinbase Developer Platform exposes blockchain connections, secure custody, and staking infrastructure originally built for Coinbase’s own products. Armstrong says seven of the largest banks are building stablecoin wallets, payments, and trading services on it. “It’s still pretty early,” but growth encourages him; he thinks it could become “kinda like AWS for us.”
  • Software becoming easier to generate also sharpens the question of defensibility. Regulatory licenses cannot yet be prompted into existence, Armstrong notes. Other assets include KYC-funded accounts, exchange liquidity, and USDC’s network effects. These are the parts of the financial business he distinguishes from increasingly reproducible software.

8. Productivity gains still have to overcome revenue lag and customer inertia

  • Senra presses on the gap between productivity and revenue. Armstrong says Coinbase has seen very material AI-driven productivity gains, but “it has not materially moved revenue yet.” One explanation is a lag, perhaps six months; another is a “red queen race,” with competitors improving equally. He favors the lag explanation, particularly because he sees less aggressive adoption in traditional finance, but calls it his guess.
  • The opportunity remains substantial on the figures discussed: fintech captures about 4% of global financial-services revenue, traditional finance 96%. Armstrong credits Nubank’s David Vélez for articulating that competitive framing. Beyond taking incumbent share, agentic activity represents a new market whose size nobody knows.
  • Coinbase is fourteen years old, yet financial habits remain stubborn. Armstrong recalls assuming as a child that a clearly better product would spread within a few years; paper rent checks illustrate how slowly behavior changes. Slight improvements rarely induce switching. A fully agent-enabled account might provide a dramatic enough difference—and “your agent might help you switch.”

9. Financial interfaces climb from orders to proactive advice

  • Armstrong’s first level is today’s explicit instruction: buy one Meta share. Level two expresses a portfolio objective—make Meta 10% of holdings, building the position over time and perhaps buying dips. The user specifies the desired outcome while an agent handles more of the execution.
  • Level three starts with a thesis rather than a security. A user might believe AI’s bottleneck is energy and data centers, or expect the “looksmaxxing” economy to grow. For the latter, Armstrong imagines a basket including Eli Lilly and Novo Nordisk as GLP-1 manufacturers, plus relevant supply-chain instruments, potentially incorporating take-profit and stop-loss instructions.
  • Level four supplies ideas when the customer has no strong thesis: an informed conversation, community ideas, or copying a trader with a strong record. Armstrong wants advisers who bring him opportunities, rather than waiting for instructions. “Everybody now has a private banker in their pocket” is his destination, extending to tax savings and automatically moving idle pay into an account earning 4%.

10. Physical execution remains the next step change

  • Asked what AI still cannot do for him, Armstrong answers humanoid robots. Software can perform knowledge work, but cannot yet travel to represent him or move things around the world. Building factories, operating mines, and running laboratories would open a much larger labor force: “That’s the next step function change.”
  • At New Limit, the biotech he co-founded, the team has discussed when to test humanoids in the lab. Armstrong says the time has not arrived. His practical threshold is a simpler task being tested elsewhere: “It can’t do barista yet, so once it can do barista, I think we’ll start testing it in the bio lab.”
Full transcript
David Senra

I wasn’t expecting to start here, but we were talking before we started recording about brainwashing ourselves into behavioral change and things we want to change about ourselves. We were talking about when I had Tobias Lütke on the show.

Brian Armstrong

Yeah.

David Senra

He said he used to be terrified of public speaking, and he would write down every day for, I think, 10 minutes that he loves public speaking. Now he loves public speaking. He was talking about how you can literally change the way you think and feel, and I was like, “Yeah, I resonate with that.”

Brian Armstrong

I think he’s right. I’ve definitely experienced this in my own life. I’ll write down something that’s a goal or something I’m trying to work on. I think some people have this kind of woo-woo thing about it, like, “Oh, it’s your—you’re manifesting something in the universe.” I think it’s actually much simpler than that. If you repeat something often enough, it’s top of mind, so during the day you start to notice opportunities that might help you move a little closer to that or practice it. It’s just a way of keeping things top of mind.

Another mental model for this would be that your brain is like a neural network, and every time you reinforce something, it updates the weights in your model, kind of like an AI model. It’s literally just repetition. People have known this for thousands of years, right? Prayer and meditation, and repetition, strengthen the neural pathway. I don’t know if that’s accurate from a neuroscience point of view, but it’s a metaphor for how I think about it.

In my twenties, I remember writing down, “I want to start a billion-dollar tech company.” I didn’t even have any business writing that down because I’d never started a million-dollar tech company or a hundred-dollar tech company. But I wrote it down a lot, and I started looking for ideas and opportunities and trying things. Some of them worked, and some of them didn’t. Then I moved to Silicon Valley when I was maybe 29, and within six or seven years, I was the founder of a billion-dollar tech company.

Early on as a CEO, I think I was pretty conflict averse. Sometimes being CEO means you have to have hard conversations with people all the time. Two people aren’t getting along, you’ve got to fire one, or you’ve got to make a hard call. Who do I want to disappoint today? Only the biggest problems come to you. Hopefully, the rest are getting delegated.

I started writing this down. At one point, I thought, “I need to reprogram my brain to not think of this as super stressful, but to think of it as getting valuable things done to move forward toward the goal.” I remember writing this down a lot: “I love conflict. I love conflict. I love conflict.” It doesn’t mean I go out and actively seek it or just piss people off, but if it does happen, I can get somewhat excited about it now. I’m like, “Okay, I’m about to do something difficult and meaningful. Let’s move it forward.”

If I’m in a meeting with the team and I’m thinking, “The work is not good enough. There’s something wrong here,” and I feel myself getting angry, I ask, “How can I use this as empowered anger to help create change that will create a better outcome?” Those are some things I’ve written down and practiced.

David Senra

Let’s say you’re in a product review meeting and you think the work isn’t high quality enough.

Brian Armstrong

Mm-hmm.

David Senra

Previously, you would suppress that anger?

Brian Armstrong

Yeah. I think I was afraid of big emotions and saying something I would regret, so I was bottling it all up inside. People could tell sometimes, too. I’d be like, “Okay, guys, I think this could be a little better.” I became more comfortable with it over time and was like, “Look, this isn’t good enough. It’s not meeting the bar. Change it by the time I meet with you tomorrow, or maybe I’m going to put—you’re not the right person to work at this company,” or whatever it is.

Sometimes afterward, people will send me notes and say, “Thank you for saying that. We could all sense that this thing wasn’t going in the right direction, and we’re fired up. We’re on it. We’ve redoubled our effort.”

David Senra

Yeah. People respond to emotion. If you’re in the meeting, you’re the leader of the company, and you’re actually angry, maybe even pounding the table, it’s like, “Oh, he cares about this. This is very serious to him.”

Brian Armstrong

Yeah.

David Senra

The way I think about this—we talked about it when I saw you in New York, I think, two months ago. You were like, “Dude, I listen to your other podcast, Founders, in the gym.” Then you and I had dinner two weeks ago, and you said, “I listen to Founders in the gym.”

Brian Armstrong

Yeah.

David Senra

There’s a hype to it. It’s almost like I’m transferring the emotion and excitement I have from this crazy life story of an entrepreneur to the listener.

Brian Armstrong

Yeah.

David Senra

One way I think about this that’s really important is the people I admire—the way they talk about their products, like Enzo Ferrari or Steve Jobs. I think I said in the Enzo Ferrari episode that he describes his products the way you talk about your lover.

Brian Armstrong

Mm-hmm.

David Senra

This is not a fucking car to him.

Brian Armstrong

Mm-hmm.

David Senra

In my own work, I’m like, “Listen, I want to be embarrassingly passionate about what I’m doing.” I think that’s the mark: Am I embarrassingly passionate? We were talking right before we started recording. These aren’t interviews; they’re conversations. People ask, “How do you pick who’s on the show?” It’s a simple question: Who am I intensely interested in talking to?

Brian Armstrong

Yeah.

David Senra

That’s it.

Brian Armstrong

I don’t know if you feel this way about your work, but I always feel like it’s not good enough. I’m constantly stressing about things in the Coinbase app. “This thing sucks. It’s too slow. Is this broken? Why doesn’t this feature work?” I’m obsessive. It stresses me out, but I use that as motivation and fuel. I’m pinging 30 or 40 people a day about how this part is moving along, why that part isn’t moving, and trying to create a sense of urgency in the company.

I don’t think I’ll ever feel like it’s done or good enough. To me, it’s still so broken based on what it could be.

David Senra

All the top performers are like this. David Ogilvy has a great line about people who create some of the greatest art, products, ads, and companies: They’re divinely discontent.

Brian Armstrong

Yeah.

David Senra

I was just on the phone with my friend Karim Atiyeh, who is the co-founder and co-CEO of Ramp. He’s like this.

Brian Armstrong

Yeah.

David Senra

Every time I’m in Miami, we spend a bunch of time together. I accidentally made a bunch of money because I was at Karim’s house for dinner. It doesn’t matter how well the company is growing or how fast, or that they’re creating all these new products. Everything is fucking broken to him. He doesn’t like anything going on in his company, which is very common because you see the real thing.

Brian Armstrong

Yeah.

David Senra

But he also has a very high bar. He doesn’t like other products or other founders. I remember hitting the table at dinner and saying, “Goddammit, just tell me who you actually like.” I thought, “If you hate everything and tell me there’s something you like…”

He told me about a founder named James who started a company called Profound.

Brian Armstrong

Okay.

David Senra

I pulled up X and thought, “All right, I’ve got to see if Karim is saying this guy’s good. Ramp uses the product, and he likes it. That’s a good signal.” I saw the guy had already followed my Founders podcast account, so I followed him back, and he sent me a DM: “Dude, I’ve been listening to your podcast forever. We’re about to close a round. Do you want to invest?”

Brian Armstrong

Yeah.

David Senra

Karim’s divine discontent accidentally made me a bunch of money.

Brian Armstrong

I love it.

David Senra

It’s one thing to be perpetually dissatisfied and never satisfied, but you can also turn that into a negative inner monologue. What’s it like for you? You’re pissed about all these things that need to be better in the Coinbase app, but what’s going on internally? Are you super negative with yourself for letting that happen, or is it more motivational, to improve?

Brian Armstrong

What is my inner self-talk? Sometimes I consciously choose what my self-talk is. In the morning when I wake up, I’m trying to get my head right for the day, do some stretching and exercise, and listen to some Founders, listen to audiobooks.

Brian Armstrong

And jumping in and starting to do some calls and work. I’m basically repeating things that I want myself to believe and feel. Even if something doesn’t go well, I’m not beating myself up about it. I’m just like, “Fuck, we gotta make this better,” and “Let’s get after it.”

I’m always thinking, “What’s next? What’s next?” Even when something goes super well, I often don’t really take a moment to celebrate. I’m just like, “Cool. That was great. What’s next? What’s next?” Somebody could say, “Well, that’s not being present in the moment and appreciating it,” in the Buddhist sense. But I also think it’s fulfilling over time to try to go do things that you want to do. I do feel a deep sense of fulfillment over time.

At any given moment, it can feel like, “Man, I got my ass kicked today,” or, “That was pretty good. A bunch of cool stuff happened.” But over the arc, I think the goal is fulfillment. It’s not happiness or something. You’re not going to get fulfillment out of doing something that’s easy. If I was just chilling all day, I don’t think I’d feel very fulfilled by that.

David Senra

Tell me the difference between fulfillment and happiness, though.

Brian Armstrong

I think of happiness as a fleeting, temporary high, but fulfillment is a deep satisfaction that can come from doing meaningful work, having meaningful relationships, and seeing things progress over the long arc. Being useful in the world is probably one of my values. I think everybody should try to contribute to civilizational progress and help our fellow people. There’s a sense of fulfillment that comes from that.

That’s a high bar, by the way. You can do it in small ways. You can cheer somebody up by saying thank you, or you could try to make a new financial system for the world that’s more fair, free, and efficient.

David Senra

Which is your mission.

Brian Armstrong

Yeah. And I think we’re one percent of the way there. It’s a massive thing. The point of the game of life is not happiness, and it’s not comfort. It should actually be difficult. It’s like a video game. If it was easy, it’d be boring.

David Senra

Mm-hmm.

Brian Armstrong

You want to continually lean into things that are outside your comfort zone to try to make progress. That’s a sense of fulfillment.

David Senra

There’s a great story that Daniel Ek, founder of Spotify, told on the very first episode I ever did of this show. Dara was offered the job as CEO of Uber, and I think originally he said, “I’m comfortable at Expedia,” or wherever he was. He and Daniel were having drinks, and Dara tells a story where Daniel put his drink down and looked at him. He said, “Since when is life about happiness? It’s about impact.”

He said, “This is one of the most important companies in the world, and you can have an impact on it.” His whole thing is, yeah, you should optimize for impact over happiness.

Going back to when you’re upset at the quality of the product, when you see all the flaws and things that can be improved, take me through how you start thinking about that. You said it’s not a negative monologue, but are you excited because you see room for improvement? What is the actual response happening inside of you?

Brian Armstrong

Some of it is a competitive thing, I would say, with other products that are out there, or thinking about how we can make it better. Some of it is that even if there were no competitors, I think I’d still be pretty motivated to do it for some reason, because my brain is very obsessed with building things and progress. It’s like growing up playing video games like Civilization and StarCraft. You collect resources, build industry, and expand to the next thing. It’s addictive to my brain. I just can’t turn it off.

I don’t really play those games now because I don’t have enough time. There are fun games like Polytopia, which Coinbase sponsored the world championship of. That was very fun. I also don’t have time to play, but these games are fun for my brain. I think those games scratch the same itch, but they’re very short-term. It’s a 20-minute game or something. In business, it’s very much the same thing, but you do it over the arc of decades or hopefully a lifetime.

David Senra

What did you learn from these video games that you applied to building Coinbase?

Brian Armstrong

Civilization is a very interesting game. I don’t know how many of your listeners have played it, but there’s a tech tree. I actually have a printout of the Civilization tech tree in my office at home because I think it’s such a fascinating concept.

When you start, Civilization starts with agriculture, writing, and pottery. At the end state, you’re doing AI, biotech, financial markets, nanoscale manufacturing, or even future tech that hasn’t been invented yet. In some ways, playing that game growing up, I thought, “I want to add a leaf to the tech tree and expand it a little bit.” Civilizational progress is just so fascinating and exciting to me.

The other fun thing about that game is there are various ways to win. It’s not so zero-sum. You can have a military victory and conquer, but you can also achieve a scientific victory or a religious victory. There are various ways—a diplomatic victory, too. So there are different ways to win.

I’ll give you one strategic example. I remember this from a particular game of Civilization I played as a kid. You’re on one continent, battling with another civilization on that continent, and that’s consuming all of your resources on that frontier. Meanwhile, you notice on another continent that one of the civilizations is starting to take over and is going to become massive while you’re engaged in this other battle. You have some technologies you’ve achieved on the tech tree. You can gift them to the weaker opponent on the other continent so that battle continues and occupies your future opponent, keeping them from getting too out of hand and large while you’re finishing this battle over here. There are so many layers to the game. I find it fun, and you can learn some simple strategies from it.

David Senra

Do you feel that with the proliferation of AI, building and running a business is becoming more like a video game?

Brian Armstrong

Kind of, yeah. You can do more of it solo. There are probably going to be these companies that are one-person billion-dollar companies. I don’t know if that’s actually technically been achieved yet. If not, we’re probably close.

Starting off as a programmer, I was doing lots of solo projects. It’s good for introvert mode, but I realized at some point that to achieve more and advance in the game of Civilization or business or whatever, I had to find teams of people, work with them, and empower them.

Brian Armstrong

That is its own whole skill set, and it can be very frustrating for a nerdy introvert who thinks, “Why do I have to communicate through this channel and get this person? If I had just done it myself, it’d be faster.” That’s the kind of thing a lot of managers struggle with initially.

I’ve made peace with that. Coinbase has about 4,000 people now, and we can always make it more efficient. We should probably talk about how AI is doing that. You can also have a team of agents working for you 24/7. Many employees, including myself, have that now at Coinbase, which is its own fascinating thing.

David Senra

Actually, let’s talk about that now. When we were hanging out a few weeks ago, I said, “Dude, I want to do an entire episode because I’ve been loving your tweets.” I thought, “This is going to sound weird for an episode, but what if I just read your tweets back to you and then you expand on them?”

The first one went viral. It got a million or two million views. It starts out, “I’ve changed how I work over the last few weeks.” I’m going to read a paragraph or two. It goes on for a little bit.

Brian Armstrong

Mm-hmm.

David Senra

The ending is really interesting. You say, “What an exciting time to push the frontier of productivity, especially as we build the economy for agents and agentic finance. We should all keep pushing that direction.”

You write, “I have one lead agent for each area of focus running all the time, and a chief of staff agent that oversees them. Each lead manages its own sub-agents that write code, test it, chase reviews, and coordinate with team members.” There’s more, and we’ll link to your tweets in the show notes. Can you talk about how you’ve changed the way you work over the last few weeks because of AI?

Brian Armstrong

When we first started using AI at Coinbase a few years ago, people used it in the normal way: you prompt it, and it gives you a response. Eventually, we shifted to giving the agent an area of responsibility or a goal. It can have a team of people or agents under it, and it actually works on that area.

Periodically, it comes back to you with important decisions. It might say, “I know you want me to optimize onboarding at Coinbase so it works better in all these countries around the world. I looked at the data and analyzed it. The 5 biggest opportunities are these. I propose starting with numbers 1, 2, and 3. Do you agree?” I say, “Yep,” and click yes. It comes back to me with ideas about what to work on, and I align it. I’m guiding it, right?

David Senra

Let me pause there.

Brian Armstrong

Yeah.

David Senra

How good are the suggestions? Do you put them at a C-player or B-player level? When you see what you’re getting, do you think, “Goddammit, this is like a genius-level employee already”?

Brian Armstrong

When you’re doing the planning, you’re typically picking one of the best frontier models, which is the most expensive.

David Senra

Yes.

Brian Armstrong

It’s investigating and coming back with higher-level ideas and tasks. As it gets subtasks, it can farm those out to lower-level models.

Most employees are using what we have set to auto. It’s choosing the model for you at this point, because I think the overhead of a human having to choose which model for which task is unnecessary.

I should mention one other thing: we’ve been using a lot of open-weight models and fine-tuning them to do specific tasks, trained on internal data. In some cases, we’ve seen them outperform the frontier models. A much smaller open-weight model that only files and drafts compliance reports at Coinbase—that’s all it knows how to do—can outperform the frontier models because it has been trained on tons of internal data. We’re getting teams of agents like that, too. Those open-weight models tend to be much faster and much cheaper.

David Senra

Before we get to what employees will have and what you’re using, let’s go back to the tweets about how you’ve changed the way you work. When you get these responses back—“Hey, Brian, this is what you want me to do. This is what I think we should do”—what’s the feeling?

Brian Armstrong

Mm-hmm.

David Senra

Are you shocked, like, “Holy shit,” or do you think, “Okay, this is like a normal employee”?

Brian Armstrong

It quickly becomes very normal. It’s like working with other people.

David Senra

But I’m trying to get at the quality of the people.

Brian Armstrong

Inside companies, a lot of tech companies use levels. They might come in at L3 or something.

David Senra

Mm-hmm.

Brian Armstrong

A VP might be L11. I think these are easily L5 or L6 employees. You can get almost an infinite team of these L5-type employees now.

David Senra

Okay, and you think that’s as high as it goes right now? You haven’t had an experience where you think, “Damn, this could be my right-hand person or a direct report running a big part of Coinbase”?

Brian Armstrong

AI isn’t exactly like a human, just as a submarine isn’t exactly like a fish. In some ways, it already far exceeds human performance. It works 24/7, gives me answers very quickly every time, and you can make infinite copies of it, or spin up as many as you want. In that sense, it already exceeds human performance. In other ways, it’s not there yet. There are definitely times when it misses the big picture or feels a little derivative of other things out there.

The pace of improvement is very impressive. A lot of people use the generic models that come off the shelf. They’ve been trained on the corpus of data out there in the world that they can get their hands on, but they aren’t necessarily tailored to your preferences and your company’s way of working.

In addition to fine-tuning our own open-weight models for narrow tasks, we’re also developing something called Coinbase Brain. It’s an internal service that gets context. There’s now a brain for every employee, every team, and every GitHub repository. As humans work on those teams or repositories, the brain accumulates and distills lessons and knowledge.

We ran an A/B test and a user study. There might be an incident, an outage, a meeting and its transcript, code reviews, or policy documents such as GAAP accounting standards. These things can all be ingested. Every night, the system goes through them, distills key lessons, and adds them to the brain. Then, when an agent works on something in that area, it puts the brain into its context and creates an output that is more aligned with what we as a company are trying to create.

Our jobs increasingly are to align the AI by editing and cultivating what’s in the brains and resolving discrepancies. Sometimes the brain has a contradiction in it. It might say, “David said this, and Brian said this 14 months ago. Which one do you want to do?” These things improve the rate at which agents can get the work right in one shot, without a ton of feedback from their human manager.

I should clarify that humans are making the final decisions and the final code reviews. AI is eliminating tasks and getting to the answer more quickly, with less nudging and fewer corrections. Humans are still very much the ones deciding the direction. Our main job now is to align the AIs working on the software factory that produces the output.

David Senra

How do you think about what the brain is? Is it the operating system for the company now?

Brian Armstrong

You could argue that it is. What is the value of a company? It’s a set of practices, people, and know-how that have accumulated over time. In some ways, the brain is formalizing and codifying that.

You still need people to implement things, or agents in these cases. I don’t think it’s ever going to be done or fully replaced. People often ask whether you could run Coinbase with 100 people. I don’t know. My guess is we’ll probably keep roughly the same number of people, but they’ll all just be able to do way more. We’ll see.

1. Building the Agentic Economy

David Senra

So, go back to saying you wanted to get into how this is going to enable the agentic economy that you think is important.

Brian Armstrong

As you start to get more and more agents working with you, the UI has to change, too. This quickly overwhelmed me once I had more and more of these chat sessions open. It was a lot of context switching, going through all these walls of text to see what was going on.

David Senra

Wait, could you take me through how many of these things you have open at one time? What’s going on?

Brian Armstrong

Right now, I probably have seven or eight lead agents working on different areas I’m focused on, and one chief of staff overseeing them. Each of the leads can spin up agents to handle subtasks. It depends on whether they’re all running, but whenever one is running, it might have three to five sub-agents.

It quickly got to a point where I couldn’t keep track of all these walls of text and what was going on. We’ve built an internal view that we’re prototyping. It’s more like a queue or an inbox of things you need to review and decide, and they’re prioritized, too. If you have a big backlog and only five minutes, you can knock out the decisions that would move the team of agents forward the most.

It might say, “Here’s a product review. Here are a couple of screenshots of this feature we just developed, and the next step is to do this.” I can say yes or no, or use voice dictation to give it a couple of bits of feedback. That’s much faster than typing. You can say, “Yes. No, no, no. Yes, do this one a little bit differently,” and start knocking out decisions, almost like in your email inbox. I find keyboard shortcuts and voice dictation faster.

The agents keep track of all these projects and decide who else on the team they should collaborate with and get approvals from. That’s reduced a lot of the cognitive overhead.

David Senra

Why does the UI have to change?

Brian Armstrong

Well, I think—

David Senra

Because you can’t just deal with all these walls of text.

Brian Armstrong

Yeah.

David Senra

So what does the UI you want to see look like?

Brian Armstrong

I think having a queued inbox that you process—a prioritized set of tasks or decisions—is a good start. Everything appears on the same screen at the same time. The buttons are in the exact same place, and you say one, two, three, or whatever. That alone is good.

I’ve also been thinking about what the next layer beyond that might be. It’s like getting even higher-level employees up that stack. You might have a VP-level agent that says, “I think we need to expand into Brazil.” You tell it, “Go do it. I’m going to check in with you every three months.” That would be another layer of the stack. That person would figure out what team to recruit, which partners to meet with, and where to sign these deals.

There are still some areas that are a little difficult to prompt into existence. Software is getting easier and easier to prompt into existence, but regulatory licenses aren’t. You can’t prompt those into existence yet. KYC-funded accounts, for example—we have to think about what the defensible parts of the business are in this world. There’s liquidity, the network effects from having the most liquid exchange in the U.S., or the network effect of USDC. Things like that. That interface has gotten more and more important.

David Senra

Who’s building the best interface for this?

Brian Armstrong

A lot of folks are trying it. You’ve seen GrokBot kind of pioneer this chief-of-staff idea with all the different agents. I haven’t seen somebody do the prioritized inbox or queue thing yet, but I’m sure other people are playing with these ideas.

I’ve been working on a little prototype, almost like an operating system for the company that all employees could jump into and view. It’s really just: What are our top priorities? What projects are contributing to them? You can drill down and ask, “What are the subteams? Who are the individual people? What tasks are they working on? Who needs to sign off on what? What’s the most important thing at any given moment? What’s the critical path? From here to there, what’s the longest set of dependencies, and who’s working on them?”

The AI could come back and say, “Here’s the biggest risk in the company. This is a critical dependency for 14 downstream things, and right now it’s understaffed, and its velocity is below the company average. Here’s a proposed solution: move David over to that team; change out the leader; change the strategy; or narrow the scope.” You could say, “I reckon number three,” and it would just do it.

You can get to a world where you could be, you know, like Steve Jobs’s “bicycle for the mind” metaphor. This allows people to have much greater productivity. As the models get better, they’ll start to be at that VP level.

I was going to say earlier, on coordination among agents: I think this is all happening within one company right now, but agents are also starting to interact with other companies. You can think of it as an individual person working by themselves on something. Now you can get to a small team of people, and hopefully they can get more done than one person. A company can get more done than a team. A country could get more done than one company. Hopefully, the whole world—humanity collectively—can get even more done than one country, because there are lots of interdependencies between countries, like trade.

How do we get an economy or teams of agents working efficiently? We’re seeing it now within one company, but as they start to interface with other countries or companies, they’re going to need their own markets and financial infrastructure.

How do you get 8 billion humans on the planet to collaboratively create an economy for the world? They need markets. They need money. They need specialists who can come together—a great security person, a great designer, a great financial analyst, a great product person—and form a team that achieves more than the sum of its parts.

That’s what we’re doing now with building agentic finance and the payment rails for it. Credit cards are playing an early role, with requests like, “Buy me these shoes” or “Book me this plane ticket.” Credit cards are important for backward compatibility, but there are also going to be stablecoin rails on crypto. I think those will be the real-time payment rail for the agentic economy.

David Senra

But if you can spin up an infinite number of agents and you actually have a functioning agentic economy, wouldn’t that economy dwarf our existing human-based economy? Or no?

Brian Armstrong

Yeah. Someone was asking me today, “What percentage of transactions in global GDP do you think will be agent-to-agent in 10 years?” I said, “Five nines”—99.999, or whatever. I think that tipping point could happen sooner than we think. But what was the follow-up question? Are you wondering what the implication of that is?

David Senra

No, just because you spent all day thinking about this. I think another trader was going to ask you about it. You’re like, “I wake up every morning thinking about agentic finance.”

Brian Armstrong

Yeah.

David Senra

I wake up every morning thinking about podcasts. A little different.

Brian Armstrong

Both doing what we should be doing.

David Senra

Yeah. I’m just trying to process what you’re saying. Every agent could spin up an infinite number of agents and copies, as much as it wants. If you have one that’s kind of free-functioning, with no human in the loop, then we can make way more transactions.

David Senra

It’s like the economy would be bigger if, instead of 8 billion people, there were a trillion people.

Brian Armstrong

Mm-hmm.

David Senra

So why wouldn’t that be the case if you can actually build what you’re talking about building?

Brian Armstrong

Yeah, I think that’s exactly right, and that’s why I’m so excited about it. I think crypto and AI really are better together in this world. You could say that what Coinbase is doing now is taking frontier technologies like crypto and AI and using them to build better financial services, ultimately in service of creating more economic freedom in the world. That’s our mission.

I think, in the same way that crypto decentralized the financial system and improved freedom, we have an obligation to make sure AI is decentralized too. The bad version, I think, would be one or two big players having too much control. They could be biased. In a decentralized world of AI, everyone has lots of small open-weight models. You control them. You can customize them as much as you want.

There’s an economy of these agents all talking to each other, which means you have choice. You have choice in the economy. Do I want to go with FedEx or UPS? You could choose different agents with different specialties and providers, and assemble them using the financial markets.

A lot of folks in financial services are a little skeptical or afraid of AI, and they’re saying, “We don’t want those AIs to touch our accounts because they might do something bad.” We are building very strict controls and permissions, and the customer always has to be in control. But to me, this is an opportunity. I’m thinking, we should lean in and be the most AI-enabled financial account and make it safe and trusted. That’s what Coinbase can do. That’s what we did with crypto.

David Senra

I was gonna say, that’s what you did with—

Brian Armstrong

Yeah.

David Senra

…Bitcoin.

Brian Armstrong

Yeah. That’s what we’re known for. We’re known as the most trusted crypto brand, and we store more crypto than any other company. So why can’t we do that with AI?

David Senra

I want to go back to this idea you said earlier. You were like, “Hey, what if you had a VP-level intelligence for AI?” And you said, “We should expand to Brazil. Just go do it.”

Brian Armstrong

Mm-hmm.

David Senra

Have you spent any time with Scott Wu from Cognition?

Brian Armstrong

A little bit, yeah.

David Senra

I’m gonna see him; he’s coming back on the show in about 2 weeks. I think he said on the show, maybe he said it off camera, that it’s just a matter of time before you have an agent that you set on a task and it just runs on that task for an entire year.

It could build a company. It could be doing a lot. What could a human, a very competent human, get done in a year? I had a weird thing come to mind. I’m listening to this audiobook where President Polk—I think, I don’t know, he was the seventh president of the United States—is doing the westward expansion. He really wants California. Mexico owns it at the time, and he sends 100 soldiers out there. “Just figure it out. Go from Washington, go take California.”

Brian Armstrong

Mm-hmm.

David Senra

It takes these guys about 2 years, but the 100 people wind up taking California.

Brian Armstrong

Mm.

David Senra

They lose it, but I thought, “Oh, there’s an analogy here.” It’s like, okay, here are 100 really competent people. I won’t know what you’re doing for 2 years. You come back, and you actually have a new state.

Brian Armstrong

Yeah. High agency. That’s one of the most important things to look for in anybody you work with, and I guess in agents.

David Senra

Yeah.

Brian Armstrong

That’s what the word “agent” means. It’s a model with agency.

David Senra

What is the thing you’ve had an agent do inside Coinbase that most surprised you?

Brian Armstrong

Okay, well, I don’t want to say the name of the vendor because it’s probably bad for them, but we have so many vendors, and vendors have their own cybersecurity issues. I like the idea of having fewer vendors and having control of our own software. If we can save money on it and have better security, let’s do that more.

David Senra

Say more about limiting the number of vendors.

Brian Armstrong

There’s been a massive increase in cybersecurity events, right?

David Senra

We just had Nikesh Arora on the podcast. He was talking about this.

Brian Armstrong

Yeah, yeah. I think the number of cyber incidents, both in crypto and traditional finance—it’s happening everywhere—has roughly doubled year over year. In the industry, they call October “Black October” because it seemed like every week there were multiple incidents. Obviously, there have been a few in the crypto space recently too.

Security is our bread and butter as the most trusted brand. We use frontier AI agents to review every line of code before it goes into production to look for vulnerabilities. We store 99% of the crypto entirely offline, disconnected from the internet. That means air-gapping it, distributing it around the world, and making it resistant to natural disasters, nuclear war, and all these kinds of things. We’ve been investing in security for a long time.

We realized at a certain point that one of our biggest risks was actually third-party vendors, because we have less control over them. We hold the vendors we work with to a pretty high standard; they have to pass security audits. But we’re looking for ways to reduce the number of vendors by rebuilding things in-house. Why take the risk?

David Senra

It seems like it’s also more fun for you. It seems like you’re having more fun now because you’re just building more stuff.

Brian Armstrong

Yeah, the cyber part of it is scary, but we’ve gotta be vigilant about it. In general, building has been lots of fun. It’s been very energizing. I almost worry about it sometimes because I’m just so excited. I can’t sleep at night. I’m like, “This is so interesting.” Then I have to go to sleep, and I wake up and it’s the first thing I think about.

We’re living in a golden age of building, productivity, and innovation. I feel very lucky to be, hopefully, in the prime of my career in this moment, because it’s really cool.

David Senra

This is something I asked Sam Altman when he was on the podcast: everybody uses AI, everybody hates AI, right? Obviously—

Brian Armstrong

I don’t hate AI, for the record.

David Senra

No, no, no. This is an audience full of entrepreneurs. If entrepreneurs aren’t using AI, they’re fucked. That’s the general population. You’re super excited; you can’t sleep, right?

Brian Armstrong

Yeah.

David Senra

These are tools, and they drastically increase how creative you can be, the stuff you can build, and the problems you can solve. Why do you think there’s such a disconnect between the way you and I feel about it and the way the average person feels?

Brian Armstrong

Most people are wired to be a bit cautious about change, right? I find evolutionary psychology very interesting. A lot of explanations of human behavior can be traced back to evolution. Probably people who were somewhat fearful of change lived longer and reproduced more, so there were more of them. Whether it was a giant flood, half your tribe disappearing, or something like that, these things would all be scary and bad from an evolutionary psychology perspective.

Luckily, there are a small percentage of people in society who, for whatever roll of the dice, genetically or psychologically or whatever, are excited enough about it, because society needs some of that to advance. But you don’t want everyone to be a crazy entrepreneur. That might be totally chaotic.

David Senra

That’d be fun.

Brian Armstrong

It’d be good for your podcast.

David Senra

Yeah.

Brian Armstrong

Yeah.

David Senra

Eight billion people listen to Founders now.

Brian Armstrong

Yeah. Yeah. Imagine that.

David Senra

The reason I brought it up with Sam Altman is that he said something on the podcast like, “There are a billion people using ChatGPT every week,” or something. It’s a crazy number.

Brian Armstrong

Yeah.

David Senra

It’s a big number. Everybody’s using AI, and everybody hates it. Outside of entrepreneurs—

Brian Armstrong

Yeah.

David Senra

—and creative people and stuff like that, it’s a very interesting disconnect to me.

Brian Armstrong

Totally.

David Senra

It doesn’t help when they go around saying they’re going to take all your jobs and kill you. But that aside, not everyone who hates AI can have heard that message. It’s confusing to me.

Brian Armstrong

Obviously, I do think their messaging could’ve been better, but there’s something deeper going on. Even if they had perfect messaging, I think people would still feel some fear and resentment about it. I remember talking to one of Amazon’s board members one time, and they were saying, “You go into a room and ask, ‘How many people are worried about Amazon, local businesses, or the workers?’” A bunch of people raise their hands. Then they ask, “How many of you used Amazon in the last month?” Everybody raises their hand.

So people can hold these two things in their head at the same time: “This is a useful thing, and I have some concerns about it.” I guess it’s that simple. There’s a great X account—I think it’s called The Pessimist’s Archive or something like that.

David Senra

I love this. Yeah.

Brian Armstrong

You can go back and look at things historically. When the bicycle came out in America, I think people were freaked out about it. They were like, “What? Bicycles?” There were all these fears that you wouldn’t know where your children were anymore. They could be anywhere in town. Licentious behavior was going rampant.

David Senra

Marc Andreessen was on the show.

Brian Armstrong

Yeah.

David Senra

You know Marc. Since you’ve been on the show twice, you know we don’t know what we’re going to talk about. I don’t send you questions. I was just prompting him because I know there’s a lot of business history in his head. He’s so good at that.

Brian Armstrong

Oh, he’s so good at that. Yeah.

David Senra

I’m embarrassingly passionate about business history, so I was just prompting him. He asked, “Do you know about bicycle face?” There were literally newspaper stories saying that when you rode a bicycle, you went so fast that the wind would spread out your face and it would get stuck like that. People genuinely thought that. The other story he told was that guys didn’t like bicycles because all the women were in the village, and it was hard to get from village to village. Now you have this girl on a bicycle going to another village and seeing other guys.

Brian Armstrong

Yeah.

David Senra

It was a hilarious story.

Brian Armstrong

You could imagine someone at that time saying this was going to disrupt the fabric of society. When I was a kid, violent video games were a big fear. My mom was scared to death because I was playing Doom on the computer.

David Senra

There were congressional hearings about this.

Brian Armstrong

Congressional hearings. When cryptography first came out, you can go back and read the articles where Congress almost banned it for commercial or civilian use because they said, “This is going to be used by terrorists.” I just don’t want someone reading my email at the coffee shop. It’s amazing how good we are as humans at scaring the hell out of ourselves about the unknown.

The precautionary principle says we should enact all these regulations in case something bad happens. But oftentimes, the thing you’re imagining in your head isn’t actually the real risk. It’s either something totally different or something much smaller, and you can accidentally kill whole industries. The Nuclear Regulatory Commission was created to protect us against, I guess, the next Chernobyl, but no new nuclear plants got built.

I have a fun list of these on my personal website. I think the precautionary principle is super dangerous. You should wait to see what happens before you put a bunch of rules in place to prevent it, because it might be something totally different from what you think. We’re all pretty bad at predicting the future. Somebody might say, “Well, Brian, what if it’s so bad that the whole world ends?” I’m just skeptical. There’s no historical precedent for that.

David Senra

Humans are usually good at solving problems when they’re presented, especially if there’s some kind of existential threat. Do you know DHH?

Brian Armstrong

Yeah.

David Senra

He was telling me that he and Toby travel together and race cars together. They’re working on Omarchy.

Brian Armstrong

Yeah. I don’t know why people in Linux or open-source projects sometimes do this, where it’s not pronounced the way it’s written and you have to pronounce it differently. I don’t know if it’s a troll or what. You’d have to ask him. But I guess it’s pronounced “Omarchy.”

David Senra

Okay. They’ll work on it in between turns racing, in the middle of a race and stuff, which I think is funny.

Brian Armstrong

Yeah.

David Senra

He was saying that he wishes—he’s trying to convince Toby, because the internal AI tools they’re building at Shopify are so good. He said, “This is better than anything on the market. Give it to me. I’ll buy it from you.”

Brian Armstrong

Yeah.

David Senra

I think they’re saying no. Right before I came here, I was on the phone with my friend Karim from Ramp. They’ve been building a ton of AI tools, all centered around problems they had inside their business. Their organizing principle is simple. It fits on a Post-it note: “Does this save a business time or money?” Ideally, both. They decided these tools were so valuable to them that they were going to make them available to the world and sell them. Which area are you in? Are you going to keep the tools yourselves, turn them into existing products for customers, or what? How do you think about this?

Brian Armstrong

We had the same debate internally on a couple of these projects, like that internal harness I just told you about with the prioritized queue. I do think it could be useful to other companies, but it’s much more difficult to build a tool that could be used in any company than one just for our own company. It’s not just twice as hard. It’s probably 10 or 50 times as much work to generalize it and make it good for everyone, then deal with every company’s security needs and different billing preferences. It’s a nontrivial amount of work.

The other thing we could do is open-source it or something, but then you start to get tons of people making requests for different things. That’s additional work. I also think a lot of these coding harnesses and AI tools will end up having open-source versions, which means every company should probably just run them internally. I don’t think that’s going to be a massive revenue driver. A huge part of the value chain might not get captured with these coding harnesses.

That’s just one example we looked at. In other cases, we built infrastructure to get our own products working. We had to connect to every single blockchain, build secure custody, and build staking infrastructure. We’ve now exposed all of that through our infrastructure product, Coinbase Developer Platform. Lots of companies are building on it. We’ve got seven of the largest banks in the world building on top of it for stablecoin wallets, payments, and trading.

David Senra

Is that going to turn into a big business for you?

Brian Armstrong

It’s still pretty early, but it’s growing really well. I think it’s going to be kind of like AWS for us. I’m pretty bullish on that product. We’ve built a lot of crypto infrastructure. We haven’t really put any of our agentic tools into it yet, but we’ll see. Maybe.

David Senra

Okay. So you’re still undecided.

Brian Armstrong

Yeah.

David Senra

Okay.

Brian Armstrong

I haven’t seen one yet that shouldn’t just be an open-source, free project.

David Senra

The reason I ask is, how do you think about Brain? Is that like an operating system? He went into detail about how he’s running the company, and they were building AI tools in 2011.

David Senra

He essentially has a proprietary AI operating system inside Bending Spoons. It’s getting better and better, and he feels like that’s why he’s able to do the acquisitions he’s doing, then improve the products and make them much more profitable. I asked him, “Would you ever sell this to anybody else?” He was like, “It’s not valuable to anybody else. It’s so fine-tuned to us.”

Brian Armstrong

Mm-hmm.

David Senra

And he said using it ourselves instead of selling it would be a distraction, because he should just buy the whole company—

Brian Armstrong

Yeah.

David Senra

—and then layer this on top of it.

Brian Armstrong

Yeah.

David Senra

It’s very similar to Alex Taubman of Long Lake, who was on the show. Same thing.

Brian Armstrong

Mm-hmm.

David Senra

They have their own internal AI operating systems. He found it better if you want to induce organizational change to just own the company you want to change. It’s a lot easier than selling the tool to other companies and not having control over how they use it.

Brian Armstrong

Yeah.

David Senra

I just thought that was interesting.

Brian Armstrong

Yeah, totally.

David Senra

Do you have any thoughts on this?

Brian Armstrong

When selling B2B software, you’ll get a very long list of customization requests. I feel like it’s taking all of our time and energy just to stay on the frontier of our own stuff.

David Senra

And the tool is better for you because it’s hyper-specialized to what you’re doing.

Brian Armstrong

Yeah.

David Senra

It knows everything about your company and what you’re trying to do in the world.

Brian Armstrong

And it can make our own products, where we already have billions in revenue, much more efficient, as opposed to trying to spin up a new product line. There’s always a balance in how many things you should be doing. There’s value in focus, but you can also get really powerful effects by having multiple pieces and being one of the only companies at scale, with economies of scale, that does all these things. At the moment, I’m trying to be more disciplined about not launching too many new products. We’ve got to focus on the ones we’ve got.

David Senra

You said something fascinating when we had dinner a few weeks ago about the fact that if productivity is increasing, why is revenue lagging? Do you want to elaborate on this?

Brian Armstrong

Yeah. I think a lot of companies are trying to figure this out right now, because we are seeing a very material increase in productivity.

David Senra

Because of AI.

Brian Armstrong

Because of AI. It depends on which firm you’re talking about, but for us, it has not materially moved revenue yet. There are 2 theories on this. One is that it’s a lagging indicator. We’re just now getting a step-function change, and it’ll show up in 6 months or something. It’s just a lagging indicator. The other theory is that because we’re getting higher productivity, but others are also getting higher productivity, it’s a Red Queen race. You have to keep running faster to stay in the same place.

I don’t think it’s the latter. There are certainly some other companies embracing and leaning into this in a big way in the fintech space, but I don’t see it happening as much in traditional finance. So I think it’s probably a lagging indicator, and we’ll see it show up soon. That’s my guess, but it is sometimes a bit of a paradox that it hasn’t moved the needle more.

David Senra

Got it. All right, I want to go back to another tweet: “We got flying cars. Now it’s time we get a modern financial system. Let me Coinbase-pill you for a minute. Financial services is a massive industry. All fintech companies combined capture about 4% of total global financial services revenues, while traditional financial services companies account for the remaining 96%.”

This was something I was shocked by. I think David Vélez from Nubank said it, and then I asked Nick from Revolut, who was also on the show, where all of fintech is something like 4 or 5%, less than 10%.

Brian Armstrong

Mm-hmm.

David Senra

Everybody thinks fintechs are competing with each other. They’re all kind of going toward this super-app idea.

Brian Armstrong

Yeah.

David Senra

David Vélez—and I think you’re making the same point—says, “Well, actually, the huge market to tap into and grab share from is traditional finance.”

Brian Armstrong

Yeah, totally. I give David a lot of credit for articulating that point well, and he’s right. The big opportunity is taking more share from traditional finance. Also, think about the TAM from the agentic economy I just mentioned, right? That’s kind of a whole new TAM that nobody knows the size of. That’s more like a blue ocean.

If we at Coinbase can be good at using these frontier technologies—crypto, AI—to update the financial system, it’s the only… Crypto are the rails. AI is the interface. If we can do that, the expansion is enormous, right? It’s less about any individual fintech that we’re competing with.

David Senra

When I heard that the first time and then read your tweet, I was like, “How could that even be true?” I feel like there have been fintechs around forever.

Brian Armstrong

Yeah.

David Senra

How old is Coinbase now?

Brian Armstrong

14 years.

David Senra

14 years. Revolut has been around for a while, and Nubank has too.

Brian Armstrong

Mm-hmm.

David Senra

There are a ton of other companies and applications in fintech. Why do you think it’s so slow, and why does traditional finance still have essentially the entire market?

Brian Armstrong

As a kid, I always assumed that if I saw something better, the whole world would be using it in a few years. It turns out the world has a lot of inertia. Maybe it comes back to this psychology point we were making earlier about people fearing change, right? If there’s something they’ve used for a long time—for example, a lot of people in America still pay rent with a paper check—it seems bizarre. But once you learn how to do something, and frankly, when you reach a certain age, people sometimes stop learning new things. They’re just creatures of habit at that point. They don’t want to change. So it sometimes takes a generational shift for these things to happen. Progress moves a lot slower than it would in my ideal world.

David Senra

There’s a founder of a very big company who’s going to come on the show soon, so I won’t say who it is. The point he made was that if you keep trying to sell to tech startups, you’re going to build a very small business. He sells only to Fortune 500 companies, and he said, “Let me tell you, Fortune 500 companies, they have a problem. When they find a solution, they think it’s done.” So he has almost no churn on his product.

Brian Armstrong

Mm-hmm.

David Senra

I mentioned James from Profound earlier. He, Kareem, and I happened to have lunch in New York 2 or 3 weeks ago, and James was saying, “It’s fucking crazy.” One day, everybody in his engineering department is using Codex. The next day, it’s Claude. The next day, it’s something else. The switching is just intense.

Brian Armstrong

Mm-hmm.

David Senra

So in the real economy, if you have a bank and you’ve been with JPMorgan or whoever, why would you even change?

Brian Armstrong

Yeah.

David Senra

Do you have to wait until these people die? What’s going on?

Brian Armstrong

Yeah. In any industry, if you make something just a little bit better, most people won’t switch. It’s got to be something dramatically better or different. Why did people stop doing Google searches and start using an LLM? It was something dramatic. It wasn’t just a slightly better Google search.

A lot of people wanted to use crypto early on. That’s fine; that’s how we got started. Now we’ve brought together the Everything Exchange. That’s a good step. But what’s going to get people to switch is, I think, the billion-plus people using AI agents now. They’re going to want to start benefiting from that in their financial accounts, right? If the legacy players aren’t going to allow that to happen and be fully featured through the agentic interface, to me, that’s a reason to switch. And by the way, your agent might help you switch.

David Senra

How does a human benefit from agentic finance stuff?

Brian Armstrong

It can just help you make more money, right? That’s a good product.

David Senra

Selling money is very—

Brian Armstrong

Yeah. I’ll give you an example. Today, if you go in to trade in a traditional brokerage or something, you have to be like, “All right, I want to buy 1 share of Meta or something.” You go to the Meta screen, type in 1 share, and hit the button. That’s not how people are going to trade. I have to know what asset I want, what type of order I want to do, and how much. That’s level 1.

Brian Armstrong

Level two is, “Okay, I want to buy Meta, but I want it to be 10% of my portfolio and leg into that position over time. Maybe when it dips, buy the dip. Help me get there over time.” That’s level two.

Level three is, “I don’t really know what asset I want to buy or when, but I have a thesis that the bottleneck for AI is going to be energy and data centers.” Another thesis might be, “I think the looksmaxxing economy is going to be massive,” like longevity drugs or whatever. You can type in a thesis about the world, and it’ll say, “All right. To get exposure to looksmaxxing, Eli Lilly and Novo Nordisk are the biggest GLP-1 manufacturers, and here are some other instruments that might be useful in the supply chain. We put together this basket of assets we think you should buy at this time, but take profit and stop loss or whatever.”

You’re starting to work your way up the stack. I think level four might be, “I don’t really have a thesis about the world that I have super high conviction in, but I want to talk to my smart friend who’s really good with money.” That’s where you’d go into your agent or see a community leaderboard of different ideas other people are pursuing and say, “Oh, that’s smart. I follow this person who I think is a really good trader and has a great track record. They just did this trade. I’ll copy trade it.”

If you just want to talk to your agent about what you should be thinking about in the world, everybody now has a private banker in their pocket, which previously only wealthy people had. When I meet with my financial advisors, they come to me with ideas. I don’t have to come to them with, “I know I want to buy Meta.” They have a whole team of people who come up with things they think I should do. So come to me and bring me the ideas. I don’t want to come to you with ideas on what to invest.

That’s how I think it’s going to go up the stack, and AI is going to help you. In a truly AI-enabled financial account, it’ll make you more money and save you money on taxes. There are all these tasks that are just a nuisance. I know my paycheck got put here, but it’s not earning anything, and if I move it over here, I get 4%. Why doesn’t it just do all that for you? That’s what AI is going to help you do.

David Senra

Mickey Malka said on this podcast that fintech put a bank in your pocket and AI will put a banker in your pocket, which is very similar to what you just said.

Brian Armstrong

I like that.

David Senra

So, in this scenario, you’re talking about what’s missing and hopefully coming in the financial space. To zoom out, what’s the thing you want AI to do for you that it’s not yet able to do?

Brian Armstrong

I think we need humanoid robots because AI can’t move things in the physical world yet, and it can’t get on a plane and represent me in another country. It’s still purely software. AI and robots—that’s how we get a truly massive labor force. Right now, it’s mostly knowledge work because it’s purely software or digital. It can do legal work and financial analysis, but once it can physically go anywhere in the world and move things, like build factories or mines or run biolabs, that’ll be the next big thing. That’s a step-function change.

At New Limit, the biotech company I co-founded, we’ve been debating when to bring humanoid robots into the lab to really test them. We don’t think the time is here yet. A lot of these humanoid robot companies are testing the barista use case: Can it make you an espresso? It can’t do barista work yet, so once it can, I think we’ll start testing it in the biolab. We’re not there yet.

David Senra

Okay, I have one more of your tweets that I want to read to you.

Brian Armstrong

Mm-hmm.

David Senra

This is related. I don’t think we ever closed the loop on the frustration that I think any entrepreneur feels: this defined discontent. They’re always upset, never satisfied, and always feel things can be better. You have a tweet that hints at this: “I’m spending a lot of my own time on Coinbase app quality right now. There are a lot of pieces that add up: instances where it’s unclear why funds are on hold, making it simpler to go through ID checks, app startup times and speed, an error that doesn’t tell you what went wrong, a button that leads to a dead end. Fixes are rolling out over the next few releases. If something in the app frustrated you recently, reply and tell me. My agents and I are reading them.”

Brian Armstrong

One of my agent leads is called Retail App Quality, and one of the ideas it came back with was to put out a post. We have all kinds of ways to gather feedback, like support transcripts and errors in production. But I also wanted to show the company a little bit that I was leading on this, because I’ve been telling people internally we need to raise the bar. I wanted to show I was contributing and throwing my hat in the ring. So it was, “All right, let’s put out an X post.” I mostly drafted it. I think the agent reviewed it, and then we put it out.

I said, “Go look at all the feedback, summarize it, bring me back the top five issues, and fire up an agent to work on them.” It put five workers on it. A couple of those are, I believe, almost live in the app now.

David Senra

How many of the things that need to be improved are coming from you compared to being found by AI?

Brian Armstrong

Collaboration is the best way to describe it. It’s like working with a thought partner or teammate. Oftentimes, I’m pushing it along, saying, “Hey, I’m still really worried about this. Can you figure out what’s going on with this team? Why has this thing not happened yet? Help me figure out if there’s something I can do to help.”

Other times, it comes to me and says, “Hey, Brian, I know you wanted me to work on app quality or make sure that our Coinbase for agents product has full coverage across the whole app. I’ve been digging through the data, and here are the biggest things I think customers would benefit from.” It comes back to me with things, and I push it on things. It’s a collaboration.

David Senra

Awesome. Thanks for doing this again, man. Always good to hang.

Brian Armstrong

Thank you.