Coinbase CTO: The Strategy to Achieve 99% AI Written Code

A year ago we were primarily writing code by hand, the old-fashioned way. Today, roughly 98-99% of our code is all written by agents. Most of our engineers are supervising 3 to 5, sometimes 10 agents throughout the day. They're producing 100x the code they were a year ago.

When you see that, it breaks assumptions across the whole stack. And soright now, an average engineer might have 3 to 5 agents working for them. In 6 months, what is unquestionably true, what I'm very confident of, we are going to have.

This is Rob Witoff, CTO at Coinbase. He helped scale Coinbase from a scrappy 2012 startup into America's largest crypto exchange, with over 4,000 global employees. And today he's revealing how modern companies should integrate AI agents in order to 100x their productivity and growth.

The cost to go from 0 to 1 is lower to build it now than it is to ask for permission to build that. And so we want everyone to have the ability to rapidly prototype and experiment with things they want to see.

Large companies historically have had a lot of layers. You get objective feedback on how you're working with people, and over the course of years you slowly get a little better and internalize feedback. And that's completely the old way to work now.

What we found agents are exceptional at. For people that are really high growth and want to grow and be the best versions of themselves they can be, and are not afraid to get the hard feedback, GPT-5.6 and Claude's 4.8+ got really exceptional in giving people the hard feedback you wouldn't otherwise hear.

People might not want to say, "Rob, stop giving me feedback. Rob, get out of my way." You know who's not afraid to say that? It's my agent. Companies that really want to grow and really want to win, I think are getting almost more productivity gains from that kind of feedback than we are from models in product development.

I think it's that transformativeright now. One of the most important things that's happened in my career, maybe my life in the last year or two, it's that.

Welcome to New Economies.

Rob, welcome to New Economies. So good to see you.

Ollie, thanks for having me. I've been excited to catch up with you here for a while.

Likewise. This is kind of a full circle moment, your first official interview since you joined Coinbase as CTO. And you know, I'd love to begin with, Coinbase started 13 years ago. Everyone has been a little bit skeptical around, you know, what is happening in crypto.

Coinbase is now the largest American crypto exchange. What do people still not realize about crypto and its biggest opportunity?

We've been here for 13 years at Coinbase, and Bitcoin's been around, or cryptocurrency has been around for 17 years now. The staying power is here. And one of the big shifts that we're navigating that we're all really excited about now is for years the killer use case of crypto was speculation, and we've been working towards what the big global scale moment would look like.

That's here now. People are using crypto rails because they're more efficient rails. They're using protocols because they're the best platform for stablecoins and large transfers and trading and new asset types and new trading instruments to be built on top of.

And so we've started to see real large-scale exponential growth all across the business. And I think we've been working towards this for so long. And now that we're seeing these new, more efficient crypto rails underpin other parts of the economy, stablecoins being one of the big growth areas, agentic commerce being another really killer use case, this is what we've worked towards for a long time.

And now we get to think about what the next chapter of scaling and really playing a bigger role in the world stage looks like.

What does the next chapter look like on that global stage?

Yeah. So there's been a lot of work happening with regulatory clarity, number one. You've probably seen the GENIUS Act and the Clarity Act. I think as the CTO, where I've spent a lot of time is internally on how are we scaling our systems, how are we making sure our quality bar is extremely high.

People want to use our products because we're the most trusted brand. That's something we think about a lot. How do we double down on trust? And for where things are scaling, one of the new really interesting interfaces that we have here is for a long time we've always built for web and mobile, and now we're building for agents internally too.

We've had new protocols like the X402 protocol we could talk about, one of the fastest growing new protocols on the internet, which is helping agents make transfers both small and large across services on the internet. So making sure that's high quality, high trust, reliable, and can scale for humans and agents are some of the areas we're spending a lot of time on now.

You know, maybe even 12 months ago, I never thought we would be allowing agents to do stablecoins for us, transactions on our behalf, buying e-commerce products. But that is now a full reality,right? We'll come back to that. You just joined CTO.

You're a couple weeks in, a month in. What is it like rejoining a company which, by the way, for the audience, I think you first started in 2014,right? Maybe just help the audience take us through, you know, how that transition went and why you started and rejoined the company.

Yeah. So I've had an amazing journey here where I got to meet Brian when he was starting the company. It's just a one-person company. And when I first came out to meet Brian, there was somebody sleeping in the office.

He was working out of a loft apartment. So that was back in 2012. And I've gotten to see that scale to about the 4,000 people we are here today. And over that period of time, I've gotten to work in just about every corner of the business, starting as an engineer, working in security and infrastructure and data and how we're recruiting and prioritizing and operating and everything internally.

And I think across that, what I've really cared about is how do we build a great organization? How do we get great people in here and build things that our customers really want to use? And having been a part of a team for so long, you get to see why decisions are made the way they were.

And one of the really interesting things about a company that scales over time is you'll often make small decisions that have big ramifications on how the company works five or ten years later. But having seen the whole journey of that, you have a really deep first principles understanding of why things are the way they are, why the culture is the way it is, and how all those things interplay with each other.

And I think that's allowed me to think really principally about what does great look like here? How do we build a good team? How do we scale our systems? How do we challenge assumptions so that we can run a really efficient and high productivity org?

And so over the last few weeks where I've gotten to start, this is maybe my sixth or seventh form of duty as CTO here, I'm able to lean harder into across the company, where are the big opportunities, how do I get people together, and how do I really get the company to the next level at all levels from how engineering works today, how we're rethinking the company with AI as that reshapes assumptions on what great looks like, the kinds of people we're hiring for, architecturally for as our customers' needs evolve as we lean harder into trading and payments and stablecoins, what are the systems we built that are not designed for that, that we need to fundamentally rethink?

And having all that context, I think, gives me the confidence to lean really hard into what the next chapter looks like.

So out of the 4,000 team members globally, how many kind of sit under, you know, technical, engineers, product, under your remit?

Yeah. So for engineering, I think around a third, maybe a little more of the company is engineeringright now. And we're actually growing that, which is pretty interestingright now in all parts of the business. And two categories we think a lot about here are as we lean harder into automation and intelligence.

One thing we believe strongly in is we need to have the best tastemakers and people with really deep domain expertise that know how to direct our agents internally, where we should be focusing our efforts. But in order to do that really well, we need a lot of great high-agency engineers that can drive rethinking, rebuilding our internal systems and partnering with those experts so we can codify that into fully automated workflows or loops or things that make the company move really quickly.

And so in doing that, we're actually growing the amount of engineers in the company quite a bit because we're trying to inject automation into every part of the business from support to compliance to security and so on. Because what we found is where there are often these legacy manual workflows, they're not just multi-step and slow and expensive.

There's also more room for human error in those too. And so where we have these well-defined workflows, similar to the self-driving car analogy, in a lot of cases in the business, we're now able to show that bringing intelligence and agents in, we can actually move faster, but also with a much higher quality bar.

And we do that by pairing these high-agency engineers with these domain experts that really deeply understand what are the problems we're trying to solve in support and compliance and going from bugs inside of our infrastructure to automatically patching code and shipping new changes internally.

That's where we're able to get this real step function increase in what we think of as our software factory internally that makes us a really high output, but also high trust modern software company.

You know, it's so interesting just listening to this. I've had a whole bunch of founders on the show recently. We were just saying the co-founder of LinkedIn, Reid Hoffman, a couple of days ago, the CEO of Webflow, the head of AI at Replit.

And the consensus is actually kind of the same. Where we are heading is all about taste and it's all about judgment. When you're looking after these 1,300 engineers, give or take, how are you coaching them about what does great taste actually look like?

And then also, what does great judgment look like?

Yeah, it's a great, great question. So taste and judgment in the modern era for engineering, I think, is akin to thinking entrepreneurially and being able to really think holistically about what are we building and why. And I think a part of that starts with we want everyone here using our products.

It's not enough to be looking at the data and sitting afar and thinking we know what our customers want. We want our people being customers, being daily users of our products internally. One of my favorite features we launched recently was you can now direct deposit directly into the Coinbase app.

And it's a small feature. It's not widely used externally, but it makes it really easy for everyone here to be a regular user of our products and get a feel for what works, what are the rough edges, where do they get trapped in some workflow in our app, and to develop their own taste for what great looks like and what they want to see.

We also, of course, augment that with we want people really understanding our customers. We do a lot of customer sessions internally to make sure people are fully understanding that workflow. We've done other things like Trading Olympics internally where we're helping encourage people to use new products.

We've been investing a lot in derivatives and futures inside of Coinbase. We're making it really easy for people to deeply empathize and develop that taste internally. I think one of the other big categories on judgment is a lot of this comes from two things.

One is just experience, having the reps and the cycles to build systems, see them scale, work with teams over time. One of the more interesting ones I think about for engineering is I think some of the best ways you develop judgment in engineering is gathering battle scars over time where you have to make mistakes to learn some of the really big things.

And one of the challenges that larger teams always have is we don't have a lot of error tolerance for our staff to make mistakes that could hit our customers externally. And so a lot of what we do then is making intentional space for people to move fast, experiment, break things, get instant feedback without impacting our customers.

And that looks like really good internal development environments, staging environments, and defense in depth so that engineers have that extra autonomy to develop that judgment once they're inside of the company too.

That's so interesting. So if you're a software engineer at Coinbase, say, if you have an idea, can I come to you and say that, hey, Rob, I've got this idea. Can I get a bunch of people to work with me and we can try it out?

I'd love to just get an insight. What does the engineering team actually look like?

Yeah, yeah. So this is such a good question. How do you go from zero to one? How do you experiment and play? It's super important that this is seamless internally. So we don't want you coming to me, coming to your lead.

If you've got an idea, we want you to go prototype it, go and build it. And one of the magical things about where we are in the world today is the cost to go from zero to one is roughly lower to build it now than it is to ask for permission to build that.

And so we want everyone to have, and this is engineers and non-engineers too, 100% of the company has the ability to rapidly prototype and experiment with things that they want to see. And I'll give you a good example of this.

We had an incident sometime last year where during the tax season, there was some tax form that I don't think we were producing in a way that it was easy for customers to use. And we did the legacy thing where people get together and start talking about how should we fix this?

What should this look like? And one of our high-agency internal non-engineers, rather than joining the conversation, used one of our rapid prototyping tools internally to mock up a fully working prototype of what he thought it should look like.

That immediately became the thing we built. And so while other people were trying to debate what should this look like, you know, how does this work? By giving people those tools, a picture is worth a thousand words, a prototype is worth a million words.

And that's how we're encouraging people to rapidly prototype, build something quickly. And we're able to get more attention and more momentum on things once people can see and touch and feel what it could look like.

You know, when I was speaking to the head of AI at Replit, my biggest takeaway was all of these software engineers, what their role looked like even 12, 24 months ago looks very different today. And my takeaway from the chat with Mikael was all of these software engineers, they're going to become AI agents, managers.

How does Coinbase think about setting up your software engineers for success when really their role has completely changed in the last, what, 24 months?

So the most important thing we think about here is the rate of change in that trend. And every three months, from prompts to agents to your harnesses and loops and graphs, every three months, there's a new thing. And every time a new frontier model comes out, it reshapes what is possible.

And we're collectively learning as an industry that in order to manage in a faster pace than engineering ever has before. So the one thing we know to be true is the rate of change is going to at least continue, if not accelerate from here.

And so our ability to predict what two, three, four years looks like from now, we can do it. We have our strategy. We have our things we're working towards, but we're probably wrong. And that's the thing you have to really lean into is rapidly embracing the rate of change and hiring people that have that neuroplasticity and are excited by adapting to that as we go.

And a couple of specific examples of what that looks like. A year ago, we were primarily writing code by hand, the old-fashioned way. Today, there's roughly 98, 99% of our code is all written by agents. And we're looking at the small exception or artisanally crafting, I like to think of it as a really small percent of the edge cases.

We wouldn't have predicted or I didn't predict last year that would have changed so quickly. Today, most of our engineers are supervising maybe three to five, sometimes 10 agents throughout the day, which is a really different muscle to exercise.

And where that's heading is going from supervising just agents to now supervising loops and full systems. And a lot of that is being distilled to what problems should we be solving in the first place? And so now we're thinking about how do we really lean into what is our product strategy, what is our architecture strategy, and who are the domain leaders that are articulating that to some model or some brain?

So we're building a lot of this into our Coinbase brain internally. And then that brain is then orchestrating our agents and loops internally that makes that a really scalable, extremely high-quality and fidelity system.

How are you setting up your agents? You know, you're stuff managing these 1,200-ish people,right? But when you actually, when you have time to actually do your job,right, and building these products, what type of agents are you using and how are you orchestrating that?

Yeah, so we've got a variety of tools internally. So the way we've broken this down is, number one, we want to be giving people the best available tooling. And so we've got access to just about every model, open weight models, frontier labs.

Those are all available. We've got a lot of harnesses available internally. So we want people to explore and experiment. But for how we're scaling that, that sort of scattershot approach only gets you so far. And so what scaling that looks like is inside of Coinbase, we've built what we call our Coinbase agent platform, or we call it CB agent for short internally.

And so CB agent is where all the pieces come together into our software factory. And so beyond just supervising an individual agent, you want your agent participating with other agents with some clear goal that's producing our product at the end of the day as one of the reasons we're all here to make the product great and futureful.

And so that all comes together through our CB agent harness where I have my agents, I have created, there's security agents, compliance agents. They get to work together to make sure when we want to ship something new or there's customer feedback coming in or we set our goals for where we want to head to, they can work together within our CB agent platform, which makes our fully integrated software factory really smooth and high fidelity across a lot of different teams that have to work together inside of Coinbase.

I kind of refer these model updates to what used to be the iPhone once every six or 12 months. Now it's literally like every few months,right? You were saying sometimes even a few weeks. How disruptive are these AI model updates to your product roadmap?

Disruptive is a good word. They are disruptive in the true nature of the word where they are not just causing some

frenetic energy internally. They are typically disrupting the way we work too. And so let me walk you through what some of the process looks like. When a new model launches externally, we try to do two things. One, we try to make that available to our staff as soon as we can internally.

But two, we're running our own internal evaluation benchmarks against that so we can figure out for our specific workflows, where do the models work really well and where are they not up for the task or where do they have some fundamentally new capability.

And the most exciting part of this for me is when a new model comes out, this doesn't happen with all model generations, but at least every second or third model generation, there's some fundamentally new unlock where it can automate a thing you weren't able to safely do in the past.

And a lot of this started earlier this year when I think it was the Opus 4.5 model came out. I think that was on February 5th of this year. And that was just an exponential unlock in software development for us.

Some of the more recent ones we've had are in general purpose reasoning where they're now very good at helping you with strategy. And as these come out, I try to look for these aha moments that I'll find personally and then share those with our peers and leaders and engineers to make sure everybody is experiencing the same type of aha moment for a new model, which lets them rethink how their internal stack and the automations they built are working.

One of the great things about that is, you know, it's really good at modifying all the tooling and automations you built to use the new model is the new model. And so we go through this cycle where the company sort of incrementally ratchets up our capabilities.

Each time a new model comes out, the challenge is ours is making sure we're sharing the aha moments so we know what those are. And we have our own custom structured frameworks for understanding what they specifically mean for a company like Coinbase.

And these are just model updates,right? We're not even talking about new models. You know, we were saying earlier we had Witoff on the show,right? And he made a really interesting point around, you know, typicallyright now there are maybe four or five AI models.

Maybe in the coming years, they're going to be 10 to 15. As we think about that, would you guys ever launch, you know, potentially your own? Have you, is that something you've thought about?

Yeah, it's something we think about and we are working on in pieces. I don't think we're going to launch our own frontier model, but what we do have is we've got 13 years of really high fidelity data on what it means to build great cryptocurrency products working with our users and all the other data sets we've collected.

And so some good examples are three really interesting examples that are a little unique to Coinbase are in compliance. We have really rigorous compliance obligations that we take really seriously and it's important to protect our customers and to protect us too.

And we've spent a lot of time using historic decisioning we've made internally to build custom internal compliance models that are helping us now increase the precision of decisions we're making and also make it much lower latency for if an account has to be reviewed because there's some compliance obligation there.

We can now do that with higher quality in sometimes minutes where it might have taken weeks or months in the past. And these are the like 99% improvements that we're seeing all across the company now. One of the next ones that's really near and dear to my heart is on how we're securing the company.

This is just one of the most important value propositions we have, being the most trusted brand. How are we securing our systems, securing our code? And one of the things I like about Coinbase is we're often a target for a lot of attackers, hackers that want to test out of vulnerability.

They'll come to Coinbase first. So we've got really good logs. We see a lot of things on the internet. We've got a lot of history for where things can go wrong. We also look at whenever there's a vulnerability we see outside in the wild.

We look at that and see how does that apply to us inside of Coinbase. So we've got a tremendous data set here. And for every change and every line of code we ship, we've always had really good scanning and auditing and reviews.

We've got the structured set of gates between code writing and going out to production. And so one of the ways we're customizing our models now is using that historical data to optimize, in some cases, open weight models where we can slightly retrain those now.

And we're thinking about how we accelerate that over time to have the best security models internally. So we always have the most secure, the most hardened systems. We've been working with frontier labs on that too to make sure we can test some of their cyber capabilities on some of the most hardened and important to harden code bases, both in Coinbase, cryptocurrency, and cryptography too.

So I think custom models there are certainly something we'll keep investing in. The last one is around agentic trading. And we recently launched a product called AI Advisor where we help people build models that can encode the trading strategies they want to build and implement those on our platform.

And again, we've got really good custom data sets there that help us customize models to give a great experience for people building on our platform. So those are just a few of the examples. And I think there will be a lot of room building on open weight models where we're pretty bullish long-term to tailor those to specific use cases to make them extremely performant.

You know, we mentioned trust earlier, and I'll be honest, anyone who's running a crypto company, sometimes I do feel a bit sorry because it is so volatile. What do you think the market still does not understand about this volatility?

Why actually is it so volatile?

So this has been a part of crypto from the very beginning where there's always these three or four-year cycles where market goes up, market goes down. And we have a saying here, Brian likes to share this, that it's never as good as it seems, it's never as bad as it seems.

And I think one of our superpowers has always been zooming out and looking at that long-term trend and continuing to build when the going gets rough or when other people get dejected elsewhere in the industry. I think one of our real superpowers is we keep powering through.

And that's why we've continued to scale and build good products and win in the long term. I think one of the shifts there, though, is human beings have, I think, a roughly fixed appetite for investing in risk-on and risk-off assets and how they're managing their finances and where they're spending their free time and entertainment time.

And I think historically that's rotated from crypto to other things. And so it's not that people's appetites for new assets or new projects is going up and down. I think it's just rotating from place to place to place over time.

And so one of the places we've been scaling the company recently is growing into what we call our everything exchange internally, where we want to provide all the things that our customers want, from prediction markets to payments to futures derivatives.

And so now that we're building out a more full featured product suite, we're seeing this stabilizing factor between the cycles because people can get more of what they want on platform. And that's something we're pretty committed to growing because we're seeing really good customer feedback from that.

You know, when you look at these super apps all around the world, I've always thought of Coinbase as the super app for financial services and the biggest opportunity,right? Where in that journey are you really truly being like this financial super app?

And where are, you know, still the opportunities to be unlocked?

I think the opportunities are everywhere for us to unlock. We're only about six months into having publicly launched our everything exchange from derivatives, futures. We started supporting equities not too long ago. But I think the real big unlock for us is we've always had this long-term thinking of how can we use, so our mission here is to increase economic freedom in the world.

And we really believe in that. We're not here to do the short-term thing. We're here to have what we believe is a really positive impact on people's individual liberties across the world. And for a lot of the things we're building to really achieve that, we have to make them run on better rails, make them more accessible, make them more open to the extent we can do that safely and in theright ways.

And so these are things like tokenizing equities and putting more of this on-chain so it's more accessible and helping people with self-custody so they have the ability to control their keys and control their crypto and things that are built on top of it.

And I think this chapter we're moving into now is, again, less of the speculative, wouldn't it be nice if, but really meeting people where they are and building things that give people real financial control, give people real economic freedom.

And so I think these next big unlocks are how do we not just support equities, but support them on modern crypto-native rails? You can do a lot more interesting things with them.

If you're enjoying today's episode, hit subscribe at the bottom of your screen. The guests on the show are getting bigger and bigger, and we need your help to grow this channel. So if you're enjoying today's guests or any previous episodes, hit subscribe, and I promise you it'll be worth it.

Now back to today's episode. You know, we said earlier 12 months ago, if we asked any agent to transact on our behalf, I don't think many of us would feel comfortable doing it,right? But now with, you know, commerce, trading, booking flights, booking hotels, it's now really becoming a reality.

You're probably the best person to ask about this. Where in the journey are we of being fully agentic with crypto payments? Are we just getting going? Are we just scratching the surface?

I think we've hardly scratched the surface. I mean, I'll give you two perspectives on this. So some of the steepest exponential growth curves we have here are on X402 transactions. And we can talk more about how that works, where it's been truly up and to theright over 100 million transactions.

And we don't see any time of that slowing down too. Personally, what's more interesting is when you're transacting in commerce on the internet, there's a lot of different sites. There's a lot of things you have to work through, and there's a lot of routine things you're doing.

If you're stocking up the house with the same thing every week or every month, if you're traversing the internet and you want to read an article on one site, you don't want to have to sign up for a subscription for everything.

You don't want to have to click through all the paywalls. You would be really happy to pay a couple cents to access a page here. You would be really happy to tell an agent to keep your house stocked up on X, Y, or Z and have a short conversation with it instead of having to navigate through different sites.

And so I think there's a lot of obvious opportunities where it's not just a nice-to-have, it's a far superior user experience for people. One of the challenges, though, which we're as familiar as anyone in crypto, is the user experience hasn't historically been very good.

It's been hard for people to use. And so where we're really excited is in seeing harnesses improve, seeing the interface become much more seamless. And a big area has been, of course, increasing the quality and model interactions. One word, for example, we hear a lot less today is hallucination.

Last year, model hallucinations were a huge problem. And so Frontier Labs have worked a lot on increasing the quality of model interactions, building in guardrails so that models are not making mistakes or doing things that are not aligned with the instructions you give them.

We're not all the way there yet, but at the direction of, with the direction of travel we have now, you can see there will be really high-quality experiences where you can tell your agent to focus on some goal.

I want to book a trip to Thailand. I want to take my family to Spain. And they can coordinate all of the steps that need to happen, do all of the research, look at all of your historical preferences, propose the full itinerary for you.

And you can give it a yes at the end, and it's all taken care of for you. That can save you orders of magnitude amounts of time compared to what that looks like today.

And it's also security,right? I mean, I think this kind of yes option is the thing that's going to make most consumers confident,right? It's not just going to go and spend $1,000 or $10,000,right, without my permission.

Yeah, and there's a lot of fundamental things we're working on here that I think are important for this to work. We're working on things, for example, like better segmentation within your account. So when an agent has access to your wallet, you don't want to have access to your whole entire wallet.

So we've got sub-portfolios and sub-accounts where you can give it access to a small amount, but making that process really seamless. And then also managing risk and guardrails and model alignment with your intent are all things we're optimizing now that I think will very quickly make the concern of misaligned models, a thing of the past.

But that's one of the limiting factors we're working through now.

I completely see this working for the consumers where if it's more the personal things,right? By me, this wardrobe or, you know, find me this itinerary to go on holiday. But when it comes to more financial services where people are a lot more risk-averse, they're cautious, obviously, with their personal capital, it's obviously a very different experience,right?

Or is it a different experience,right? How do you see the experience working for more financial service-related agents that are scouting new opportunities for consumers?

So I think there's a continuum here that holds true across really any use case. And one of the first important areas to now is getting your agents read-only access to all of your data so they can't take any action at all.

But if you could still get really crisp insight into how is your portfolio allocated? As prices change, how does your allocation change over time? How does that match your risk tolerance? As assumptions are changing, are you still managing your portfolio the way you'd want to?

I think there's still extremely safe, read-only, high-value places where we want agents to start. Where that gets into the more nuanced territory is when your agents can start to takeright actions for you and when they can make a decision or execute a trade on your behalf.

That's where we need a few things. Number one, we need theright hard constraints. So you don't want to be able to move funds that are not allocated for it. Number two, we build a lot of this internally. Escalation to human has to be a core workflow here.

If you're in, so models often work through a lot of assumptions, and that's to give you a good user experience so they're not asking you 20 questions to respond to a simple prompt. But if you're making a lot of assumptions like that for a financial application, it's not always appropriate.

And so we want agents to be bringing humans in at theright point inside of a workflow so that we're never making an assumption that you, the human, wouldn't make. Now, that has to compound over time too. And this is where the data becomes really interesting.

If you're providing the same feedback multiple times and as you spend more time with the agent, we want it to learn from you. So we think about this like a self-improving loop where you have a loop where you might kick off a trade or some financial rebalance or want to surface new opportunities for new assets that are launching.

You're going to interact with that. The model learns from you, and that gets more tailored over time with the brain we build around yourself, your interactions, the context you provide that a model has access to to make sure it's escalating to you at theright time, but not over-including you for things that become just high friction over time.

You know, as you say, we're just getting going with this. If we look in 12 months' time, where's this going to be?

This is going to surprise a lot of people where we have a tendency to underappreciate how quickly and how far exponential growth curves can change. We are unquestionably in the middle of one. One of the trends we see internally is AI has impacted different roles within Coinbase in different ways.

Some roles are impacted a little bit. Other roles are truly impacted 100X. And we see this across engineering where some of our most prolific engineers working on projects that are highly suitable for automation, like a big migration or working across many codebases, they're producing 100X the code they were a year ago today.

And when you see these 100X to, I think we'll see 1,000X in not too long from now, when you see that, it breaks assumptions across the whole stack. And soright now, an average engineer might have three to five agents working for them.

In six months, you might have loops with 100 agents working for you across the company. That breaks for us internally, our software development stack, our tools we use to ship code out, how we're monitoring things, how we're reviewing and improving things.

And there's just one limiting factor after another. But I think what is unquestionably true, what I'm very confident of, we are going to have an order of magnitude more agents running for everyone. And the mental energy we spend managing the agents, not just executing them, is going to become an appreciable part of how we're all spending our mental capacity.

And this is why it's so exciting,right? The next opportunity is us becoming agent managers. And then once we're running 5 or 10, are we going to have agent managers of agent managers,right? It's like a flywheel effect. And then, you know, what comes next?

This is where there's this constant stepping back to reassess what's the current limiting factor and working onto the next one. And this is why it's so important the people you're hiring and the people that are getting ahead are really flexible to constantly step back and say, OK, I've been working this way for a month or two.

I've noticed I'm starting to do roughly the same thing. I think a new or the latest model family can probably do this thing that I didn't trust the last family with. How do I bring them together and step back and do the next thing?

And this is just a constant scaling processright now that for people that are interested in that continuous improvement, I think it's one of the most exciting times of all of our careers.

And also just the opportunity for every consumer to get educated on what other financial products are out there,right? One thing I've been super surprised about, to be honest, is these prediction markets. Obviously, there are now more players in this space and the market is heating up.

If you asked me 12, 24 months ago, I was like, these things are just another, essentially, you know, some people will say it's a gambling side,right? But actually, when you look into it, it's such an interesting use case.

How is Coinbase thinking about this prediction markets opportunity? So I know you guys have just entered the space.

Yeah, it's an extremely interesting market where it allows you to create almost, and as someone in news media, you think a lot about this, it's becoming a more efficient way to share the news and share what's happening in the world and share the real ground truth, which is something I think the world has a real appetite for.

So I think once you get past the superficial use case of what's the weather going to be tomorrow and more what's geopolitically going to happen here, what's going to happen with technology in the next year, what's the new model family going to achieve on your latest evals or benchmarks, it becomes a more efficient way for you to navigate and understand the world.

So that's something we're really excited about. And that's really aligned with our mission to advance economic freedom for the whole world, to give more people more transparency into that. I do think for what prediction markets look like, though, we see that as one of many different order types or market types that we want to give our customers exposure to.

And so if that's what our customers want and it's something we think is providing real value, we want to give people the best experience to do that in one place. So whether it's crypto or equities or prediction markets, we want to have the full suite.

And we want to make sure those are built on modern and accessible rails so we can expose them the best way, not just the legacy way that it's worked for a long time.

Super cool. OK, so we've had agent managers. We've got prediction markets. And the third, which has also caught me by surprise, unsurprisingly, is stablecoins. Is this the next big opportunity?

Unquestionably, yes. So number one, if your agents are out there performing some tasks for you, they need to coordinate with other agents. They need to access data, different data sets, pull from the news, pull different models down, have other services manipulate images or content or write code for you, or even purchasing more inference for you.

What's the most efficient way for an agent to make a transaction? It's using some kind of digital trustless rail that's extremely efficient and can support low-cost payments. That is the killer use case or one of the killer use cases for stablecoins.

The other thing we see is for non-custodial wallets, for payments, for transactions. It's a more efficient, lower-cost, faster settling. It's a fundamentally better protocol. And so in so many of these cases, people are not using it because they think it will be great.

They're using it because it solves a real problem for them today. And that's why we're seeing such exponential growth here. I think at the same time, stablecoins are still hard to use. You still have to figure out where do you acquire them?

Where do you store them if you're going to top up a wallet for your agent? How does that work? Do you have a wallet built into your default harness? Probably not today. And so when you see these exponential growth curves and then you look at what does the user experience look like and you see all the friction that's out thereright now and all the gaps in the market, that's where you see this perfect storm of continued growth.

So I think we're at the very beginning of what that's going to look like. And there's a long way for this to run.

I think it's going to be really interesting to see are these stablecoins going to remain private or are we going to see country-owned stablecoins? Obviously, there are pros and cons to both.

Yeah, there's huge questions to be answered here. The industry is moving at a mile a minute. We're obviously having a lot of these conversations, working with our partners on this. But it's another really fast-to-navigate environment. And what's underpinning this, again, is we're solving real problems.

We're building something people want. And I think that's the best position to be in. I think a lot of what I think about is how do you win or how do you provide the best services in a market that's moving really fast?

And there's a lot of competing trade-offs there. I think you win by building the most efficient machine to create the machines. You win by having the highest velocity, highest quality software development factory. And so we've spent so much time making this streamlined internally.

And so where we're able to build great products really quickly and you see this exponential growth, what I don't think the world's fully appreciated is how quickly large companies, great products are going to be evolving with these user trends.

And where in the past, you might be expecting a flagship new release or getting a new iPhone or a new computer once every year or two, now we're starting to get big upgrades every 24 hours, every day in some cases.

And so the velocity at which the world writ large is going to be running and is starting to run is not something I think anyone really internalizes the impact of that. But it's going to be one of the wild things we watch in the next couple of years.

You know, this rebirth of even creating the new financial infrastructure,right, the new financial ecosystem. As we're having this chat, you know, I was thinking, so I'm 28 now,right? Even when I was 14, 15, I remember going, you know, with my mom and dad to the ATM with my little debit card, withdrawing, you know, $20, $30 or something.

Back then, there was nothing like this. Now we have everything available. It really is truly wild how things are growing. And if we think GBT only launched November the 30th of something 2021, what does the next even 12 months look like?

We can't think five years ahead. We can only think in like 6 or 12 months,right? How do you guys think about the velocity of change and just preparing for the next huge wave of growth available to everyone?

This is a really hard problem because this becomes a human coordination problem at scale too. One of, say, a few of the things we've done that have had really outsized gains are number one, we've historically needed, so large companies historically have had a lot of layers.

You get a person, reports to a manager, a director, a senior director, and all the way up. And it's really good for coordinating because you can set expectations at each level. But it's really slow for executing because you've got to deal with all this human coordination.

And so we made a concerted effort over the last year to flatten our company so that there's fewer layers between decisions being made and the person executing on them. And one of my heuristics for this is when we're working on a big engineering problem or we're embarking on a new big architectural upgrade, I want to have the engineer in the room that's executing it when we're discussing it and making a decision so we can completely flatten that and eliminate the games of telephone that have historically made it really hard to coordinate and execute at scale.

So that's one really interesting one. Another one is in how we're getting really objective feedback on our execution. And this is one of the most important things that's happened in my career, maybe my life in the last year or two.

It's that especially as a remote company like Coinbase, almost all of our interactions are over Slack or they're like this on a call where you can record or transcribe. And historically, you get objective feedback on how you're working with people.

You do a big performance review process. Once every six months, you get some feedback. People are maybe afraid to say what they're really thinking. And over the course of years, you slowly get a little better and internalize feedback.

And that's completely the old way to work now. That's not at all how I think a great modern organization runs. So for people that are really high growth and want to grow and be the best versions of themselves they can be and are not afraid to get the hard feedback, what we found agents are exceptional at, especially the most recent version models, the GBT 5.6 and Claude's 4.8 plus families of models, got really exceptional in giving people the hard feedback you wouldn't otherwise hear.

And so this looks like we interact over the course of the day, the week. Every Friday at 9:00 a.m., everybody at Coinbase gets a message, which is a full assessment on everything you shared over Slack and if you transcribe meetings, things like that.

What are you doing well? What are your gaps? Are you not communicating well? Are you too frenetic? Are you not setting a clear strategy? Are you not communicating well with your peers? Is there some friction? You're not picking up, but somebody else is maybe seeing.

And then my favorite section here is we always have a spicy feedback section, which sometimes is a little extreme, but it's a good prompt. And it's what is the really hard feedback you need to hear? And so what I found personally using this, and we've got many thousands of people that are using this every week now, is that things that would have otherwise taken me six months to figure out, I'm now iterating on a weekly basis.

And one of the things that I get dinged on often here is I often get too far into the weeds. I'll be too hands-on somewhere. And people might not want to say, Rob, stop giving me feedback. Rob, get out of my way.

But you know who's not afraid to say that is my agent. It's really good at saying, you're way too in the weeds here. You got to stop doing this. There's something over here that really needs your attention. Go do that.

And so we were able to get this kind of roughly real-time direct spicy feedback to people at the company. Companies that really want to grow, really want to win, I think are getting almost more productivity gains from that kind of feedback than we are from models in product development.

I think it's that transformativeright now.

So fascinating. This spicy take agent, what is the last piece of feedback it gave you that you're not doing well at?

The last piece was there's a part of engineering that I own where the strategy is not clear enough. And so people are working, you know, not in as well aligned as they should be. And so my feedback is I got to go in and fix that.

And I've got time set up next week now. And I will go in and fix exactly that. And just keeping you focused and spotting those things where everybody has blind spots is one of the more exciting times to be able to work and get feedback like that.

And the tool you're using, is this a built-in house tool or you use a third party?

This is an entirely custom tool. And we built it entirely custom for two reasons. Number one, the cost to build products is so lowright now, there are really low barriers to go and do that. Number two, for these to be really great, you got to tailor them to your culture, your values, and enrich them with as much of your personal data as you can.

And so this has been an iterative process for us over the last year. But we've tuned this to the point where I think everyone here is getting the best feedback of their careers.

This is so fascinating. I ask this question on all our episodes. If I was to come and join you as a chief of staff for a week, what would we be doing and how does your day look like?

If you were to be my chief of staff for the week, so I try to break up my time into engineering, running our shared services internally, and how we're advancing AI for the company. We would be looking at each of those independently.

What are the key metrics we want to drive there? And really digging into those to make sure we understand what's happening across the business. And then I spend a lot of time on deep dives across all three to figure out where are the limiting factors on each.

So you'd be helping me with that. And then getting together with the people so we can get in the room with the engineers, with the leads, debug what's going on, and help fix the limiting factors so we can move on to the next one.

And I try to lean in pretty hard there because people are often afraid to reach out for help. They want to work through it. But sometimes the fastest course of action is to just get theright people together, pick one path, and stick with that path instead of letting people do the painful thing over the course of weeks and months to slowly evolve.

Sometimes making a decision is much better than continuing to linger or making a suboptimal decision.

And with these 1,200, 1,300 software engineers working on do alongside you, do you do like an update once a week or hands-on? How's that kind of like structures?

As a whole org, we get together once a month. And I get together with these different groups every week, though, and we'll dive into what some of the key factors are. One of the pieces that actually two of the pieces we're piloting now, which are pretty cool, is as I've built out my personal knowledge base over the last year, and it's got all of my preferences and interactions, and it sees all of my private contacts, I built out my personal brain internally.

And so I've been ramping up people that have access to talk to my personal brain, which has all of my preferences in it. And so I've got about a dozen people with this accessright now. And they can talk to that anytime they want.

And we've tuned this to the point where I'll usually review what the brain has shared with them every couple of days, make sure it's not saying something wild. But it's usually able to give real-time, highly targeted feedback the same way that I would in seconds where I wouldn't be able to do that at scale.

And so we're going to scale this up quite a bit internally. The second thing is building on that Coinbase coach we have internally. I'm now sharing my perspective on how people are doing. I'll share that with about 50 people in my org next week.

So they'll be getting direct feedback from me at scale in a way that you would normally never do this. You wouldn't have the capacity for that. But because I can pull together all of my interactions and have these latest generation models draft this with me, I'm able to drive, I think, a much more aligned, much more cohesive org than we have in the past.

It's so interesting. I mean, all of this is just changing so quickly,right? You know, the conversations we've been having with guests on the show, I think as an overall agreement, we think since Jan of this year, this has really been the inflection point for everyone.

Yeah, I would add this has been the latest inflection point. And we're just waiting for the next inflection point for this to continue to go.

It is true. When is that going to happen? Is this going to happen in the next 12 months, do we think?

So the big inflection point this year was February 5th, where one of the latest generation models was released. I think it was Claude 4.5 and GPT-5.4, maybe it was. I think the latest set of models, GPT-5.6, felt like a really big inflection point for us and how we're using intelligence in our strategy and how we interact with people internally.

I think the next big inflection point is probably in the coming weeks when the next set of models come out, where models are starting to get really good at orchestrating across domains and projects and applications. And as full computer use becomes something that models are really tailored for, rather than having to connect an agent through MCP and onboard specific apps to make sure the API is available, your agent has access to a whole computer.

And so as long as a human can do a thing, your agent can do a thing too. And we'll take this through the same maturation arc we talked about for financial accounts, where initially it'll just be read-only. You're not going to trust it with full access.

You're going to have dedicated agent accounts internally. So just like your humans have an email and a Slack handle, agents now have the same thing. And then selectively, you'll give them more and more control over time until eventually they've got full peer access to another person inside.

And they're performing a full role, which I think puts humans in this increasingly high leverage position where now you're orchestrating more of them, you're managing more, and your ability to ship products at massive scale with massive quality and trust continues to scale with it.

So we have these kind of compounding exponential factors on top of each other, which is going to make the next couple of years feel really sci-fi.

And I think even the next few months, next few years, you know, as we will have these agent managers, we can actually go on holiday. We can actually have time off and our agents can do the work for us.

When your agents are doing the work for us, how do you spend your time outside of all this kind of AI chaos? What would you like to do in your spare time?

I'll give you two or three answers. So I've got a young child at home. So we had our first kid about 18 months ago. So that's one of my favorite things is to play with our little baby at home.

I think next to that, what I've been so passionate about for so long is that kind of personal growth and really objective reflection on how are you doing? How can you be improving? And so the killer app that I've always been hacking on in the background is how are you tracking your health and your wearable sensors?

And how does that tie into your exercise, your sleep? And so what I've been able to do recently is I pull all of that into a model that gives me weekly feedback on how I'm doing and tips on where I should be exercising more or changing my schedule or changing my diet and nutrition.

And I am just an eternal geek for ingesting more data in there and optimizing more of my life.

With the same, I mean, I'm trying to get more than four hours deep a night, but when you're busy building, you know, it's a little hard. But I try and get my 10, 15,000 steps in a day.

You've got to make some trade-offs, but sleep is everything. One of the things my agent tracks is how my sleep and my heart rate as a proxy for stress levels impact the quality of the advice I'm able to give people and the quality of work that I do.

So I have an agent that runs on top of our CB coach that augments that with my smart ring data that pulls all that together and gives me some extra feedback on top of that. And what I've found is my sleep is a really good predictor of how well I show up and how well I'm able to interact with people here.

Yeah. And then this spicy take agent you have, which gives you feedback, hopefully is always going to say, Rob is on great form these days.

That'sright.

Rob, this has been amazing. Thank you so much. Congrats again on joining Coinbase and come back in 12 months and let's see where things are at.

Thanks, Ollie. Really fun to chat. I'm sure things will be wildly different, but it's the ride of our livesright now.

I agree. Cheers.