Great. Hi, everyone. Thank you all for coming or joining and listening in on On the Block number two. We had Jack Dorsey on a couple weeks ago, and I'm really excited to talk to Brad Axen. Brad is one of the small group of people who has built Goose and has built a lot of our AI tooling at Block that I know I've had a lot of conversations with folks. I think it's incredibly differentiated, and it's been really fun for me personally to start building on this stuff a lot more and hacking together probably mediocre demos, but then get it in the hands of people that actually can do stuff with it.
For anyone who was wanting to know how Goose got its name, we'll probably answer that, but maybe do that a little bit later. Brad, I'd love to start on just your background and, you know, where you've been in your career, how you got into computer science generally, and just, like, what you spent your time doing both maybe pre and your time at Block.
Right. Yeah. Yeah, thank you for the intro, too. Excited to talk about a lot of this stuff. My background, I actually originally did most of my, like, serious programming in academia. I started in physics and got a PhD, worked on particle physics and worked at the, like, Large Hadron Collider, if y'all have heard about that. You don't necessarily think of that as a programming heavy place, but it really is. It's like extremely high scale data analysis. We kind of built a lot of novel systems to handle the data that we were collecting back then. It bridged really smoothly into machine learning, which was how I came into tech.
My first job at Block was doing growth machine learning, and then I worked on lending. The last two or three years, I've switched to working more or less full time on AI and applied AI.
What was that transition like, what do you observe in AI that made you wanna pivot a bit into going from lending and machine learning and just, like, building that capability set for Block into more full-time, like, developing AI tooling and just everything that you've done there? Like, what led to that transition in your career?
For me, it was definitely just seeing the potential of the new technology, right? I think, even though this was, like, in the early days where there was a lot more skepticism, I feel like no one is very skeptical anymore, which makes sense. Back then it was a little bit less obvious what it could do, but we were seeing it be very useful for people to help automate small parts of their job. That plus the growth curve of, like, oh, every six months we were getting a dramatic performance improvement just made it so obvious to me that this was gonna change things and made it really exciting to work on.
Right from the get-go, I think the main thing that I found motivating was figuring out how to take the current state-of-the-art models and then actually apply them into doing useful things, more than just having a conversation with it, actually using it for day-to-day work.
Like, what have you been up to then in the last two to three years? I know you kind of, like, applied science or something maybe, but taking the foundation models and just everything that exists there and making them useful. Just talk about, like, what you spend your time doing.
Yeah. The underlying theme is building agents. I often say I work in applied AI, and what I mean by that, again, is like taking those foundational LLMs and connecting them into systems that let them actually take actions for people and be really useful day to day. I work on that both like we were, we already mentioned Goose, so I was the original author of Goose, and that's a productivity tool that was originally for developers, like software engineers, but increasingly for the entire company, and actually I think we've done that for a long time. We use these agents, and we use them to speed ourselves up. I also work throughout our AI products.
For example, I think probably a lot of people are familiar with Moneybot, which is fairly generally available at this point, and then Managerb ot, which is in beta. Those are also examples of taking those foundational LLMs and applying them to real world problems. For Moneybot it's like applying it to your personal finances, and Manager bot it's applying it to run a business.
Yeah. I definitely wanna talk a little bit about Moneybot and Managerb ot and builderb ot, which I actually started spending a lot more time with this morning for the first time, and it was really cool. I've shown it to a few folks, like, on Slack, but I was actually building something today, and it was a lot of fun. Maybe before getting into some of those specific, like, application endpoints for AI, can we talk just a little bit about, like, the AI tech stack? I'm sure everyone is familiar with, you know, OpenAI and Anthropic, at least, like at a conceptual level.
Just talk about, like, what the tech stack looks like in terms of, like, foundation models as its own separate entity, and then, like, agents on top of them, and how those two things differ from each other and how they interrelate.
Right. I like the foundational models in a way because even though there's this incredible amount of complexity in what they can do and how intelligent they are, they're a really simple concept. The foundational models really just take text in and then give you text out. I think the insight that came with ChatGPT was that you could turn that, like, low-level tool into a conversation by, you know, kind of teaching it that instead of just continuing with whatever the person said to it was gonna be replying to what the person said, like it's a conversation. That I think was that first, like, layer on top of the foundational LLMs that really changed it and made it something you could interact with.
Sometime around 2024, everyone started realizing that the models were getting good enough that the output that they create in text could be code. Agents is kind of that idea of having a conversation, taking that one step further, and asking the model to write some code for you, and then the agent actually executes it. That's like the foundational stack that we're talking about. We've since like kind of in the industry called this tool calling. That's the foundation for agents and is what gives it the arms and legs to go interact with the world.
You kind of ask the foundational LLM to write small pieces of code, and then you set it up in a system where that code is gonna get executed, and then it goes and does things.
Mm-hmm.
With something like Goose, that means helping you build software. With something like Moneybot, it could mean analyzing your transaction data and giving you some insights about where your money is going. Let me just say a little bit about what I think is the current model generations are incredible. There are several that are fantastic. Like you mentioned, Anthropic and OpenAI, like Opus 4.6 and GPT-4o are both extremely powerful models that take this to like another step where those agents can really do a large fraction of what a human can do. That's also true for a lot of the open source models, like GLM 5 is super impressive. Increasingly these are kind of getting more and more interchangeable.
In the early days, you would maybe have to have a customized agent for every model. Now we're finding that you can easily switch between them and get really high quality outputs. Yeah, I can say more about that, but let me stop there.
Yeah
In case if you have more questions.
No, no. I mean, I think, like, I wanna spend time on the foundation models and it—I kind of been describing to folks as like a zero to one, which is probably not accurate, or like a big bang in December, January timeframe of Claude Opus 4.6 and relative to what came before that. Before I do that, I wanna go back in time a little bit. You said you started talking about-
Yeah
- or you started developing Goose and agentic interfaces in early 2024. It seems like I think we had a really early perspective relative to industry that we'd be able to differentiate on agentic interface and that if you think about, as you were saying earlier, as agents as the arms and legs, if you're gonna actually execute work in an enterprise environment, now increasingly even in a personal environment with Claude Bot and things, that agents were gonna be the way in which you would do that. What drove that perspective that these foundation models are fantastic, but what's actually gonna enable you to execute the work as agentic infrastructure? What drove that perspective and why did we start building this in 2024?
Yeah, in early 2024, I may be getting the timeline wrong, but sometime around 2024, we were like everyone was talking about how these models could write code, and there was a lot of truth to it. It was like in a way more compelling equivalent to like Google Search, right? Like where you're like, "I don't know how to do this. Can the model give me like a little snippet that I could use?" But it was all conversational, like you were just asking it for examples or tips. Those earlier models were impressive in many ways because they kinda knew a little bit about everything.
Still true, but now they're very, very effective at writing code, and back then it was a little bit smaller scale. I at the time we like, a lot of us found that experience more compelling than some of the tech at the time that was popular, like line completions, which was also AI driven, but it was, you were in your editor, you started typing, and it would try to like finish the line for you. We found it a lot more intuitive to learn and move faster by asking for like large, larger scale examples. The problem there was that it was actually just a really inconvenient workflow.
You were like, it didn't do anything for you. It just told you how to do things yourself, and you were like transcribing what it did. A lot of that early, or those early days work were literally just trying to get the model to say things in a consistent way so that you could detect that it was trying to give you code that was helpful, and then you would like automatically figure out how to run that or put it into your files. Around that time, they actually like OpenAI put that into their APIs as like a tool calling spec so that you could actually like more directly get that kind of output. Even back then it wasn't super consistent.
We did all of this work to try to get a consistent workflow where you could just ask the agent to do work and walk away and come back and like see what it had done. I think the main motivation behind that was just trying to actually use the tools and those like state-of-the-art models to go interact with what we were doing internally. That was the first motivation for sure. It was a bit of a mix of like literally just being enthusiastic about what that could do for us as developers, and then also thinking about that being a testing ground for how could we apply this technology for our customers too.
We've kind of circled around it a little bit, but all of this early insight and like your perspective on this and other folks' perspectives like led to us creating Goose.
Mm-hmm.
It's like us, like the royal we, like I'm on the sports team, like when I say like the Celtics, like Celtics won, but us meaning you and a few other folks that like actually developed Goose. Maybe just like taking a step back, what is Goose? Like how did Goose come out? How did the idea originate? I know you talked a little bit about that in the prior answer, but I'd love to just understand.
Yeah
What Goose is.
Yeah. It's back to that, when we were talking about how the LLMs work and then the agent being the thing that executed the code it wrote. Goose is that second half. If you're looking around the industry, I think the standard terminology we're using these days is calling it an agent harness. It's kind of the system that asks the LLM to take actions based on what the human is looking to do.
Uh-huh.
Going and reaching out to all of the different tools that it needs to take those actions. Goose is an agent harness. It's open source. I think we were one of the earliest. We had a desktop interface that was a little bit more approachable for non-engineers early. If you haven't tried Goose, the closest equivalent today is Cowork from Anthropic. I think people are really excited about that lately. Goose, we had that in February of last year.
That's something that our employees have been using for a really long time. The reason why I bring that up as another interface is because that core idea in Goose of it taking actions for you definitely started with developers in mind, like writing code, editing files, and helping you build software. Increasingly it's something that everybody is using. At Block, I think it's, as you were saying before, like pretty much everybody at the company uses AI all the time to just do operational tasks, like interacting with your Google Docs or helping you build a spreadsheet, summarizing Slack, catching up on your email. We've been doing that for a long time. Goose is general purpose in that sense. It can help you with all kinds of tasks.
Of course, me, as a software developer, I use it to build things, and that's also a really cool thing I've been seeing. You mentioned you were trying out builderb ot this morning. We'll talk more about that.
Yeah
The other incredible thing that we see with tools like Goose is that other job profiles at the company are now building prototypes, and it's this more like efficient and fluid way of talking about what we want to do for our customers. It's like, "Oh, let me show you," rather than just going through a long planning process.
That's been like really. It's been amazing for me as someone who's not technical, but I've been in the industry for, you know, 12, 15 years now and have a lot of ideas. Like, "Oh, maybe we should do this." It was. If I go back to the beginning of 2024, it was very, like me verbalizing what I wanted to do, and it was kind of hard to describe. Then I'll never forget, about, I think it was May of last year, I was at an investor conference with Jack, and I was talking about Goose and like just AI tooling in general. I was like, it was like I was having a hard time like figuring it out or like actually getting it to execute stuff for me. He said, "Don't ask it how to do something.
Tell it to do the thing that you want done, and it will do it." That was kind of like a light bulb moment. It's like I was basically using it like the early LLM models that you were talking about, where I was like asking like, "How do I, how would I do this?" He's like, "Just don't, you don't have to do that. Just tell it that you want this thing done, and it will do it.
Yeah, that's right. What's been really fun watching this space go, right, is that we kind of advance on two frontiers. One is like making the agent better, which is the thing I work on all day, and the other is that the LLMs just keep getting much better. Very excited for the like continued progress there. It, like to your point about like you just to go ask it to do something for you, middle of last year that meant it might be able to like build you a chart or two or you know like go build you like a single page website.
Now you can do that same thing and you're like, "Show me a demo of what our product would look like if we added these seven new features.
Right.
That is like the level of capability. Again, all you have to do is know what you're looking for and explain to it like what you want to see. They're all now extremely tenacious in like solving those problems for you, which I think is really cool. Then there's obviously a lot we could talk about in terms of like making reliable production-ready software, which is where I spend a lot of my time these days. In the prototyping world, I think it's just super compelling, and pretty much anyone now can build great prototypes.
Yeah, absolutely. You mentioned this term earlier with tool calling, and like the conversations I've had with you and other folks, talking about like agentic systems is basically like really good tool calling systems.
Yeah.
Can you just talk about what that or explain what that means to people and just like how, like just help us understand that concept?
Yeah. This is my whole life now, by the way. I'll try to keep it high level, but we can dig in as much as you want. There's two aspects to this. One is literally just coming up with the tools and enabling them, and that is a real art in some ways too. Like what are the tools that the model needs access to do what you want it to do? In probably most of the agents that people are familiar with, like if you've used open source Goose or Claude Code or Codex or AMP, any of these things, there's this like core kit that we all have come to expect of it, editing files and running commands on your computer.
That's like incredible for software development, and it goes a really long way. When you want it to start interacting with the rest of what you do at work, you have to go enable, like design what tools will be effective for it to solve other kinds of problems. Some of that's just simple stuff like, oh yeah, it needs to be able to read Google Docs-
Mm-hmm
-or edit a Google spreadsheet. That like opens up a whole new world for people where it can like interact in the finance space. Another example of that is, like recently I think a thing that has changed how I think about working with AI is that we've put them in Slack and then given them access to just read our code. Obviously, from Slack you don't want it to be like editing code, but in Slack it can read code. Suddenly now at any point when you're having a technical conversation with people in Slack, you can just ask the bot to like read through all of our existing code bases and explain how something works today or how we could change it to do something new.
Those are the kinds of things that we work on when we talk about really a high quality tool calling. It's just like what is it connected to and what does that enable for the people who are working with the agent. The second half of that that's critical and way lower level is how do you make those tools performant for the LLM. There's a lot of stuff that goes into that, and it's things like you'll hear terms like context engineering where we talk about like making sure you precisely control what the LLM sees so that it knows what it needs to know but it doesn't take up too much space. Like a lot of work like that that goes into the tool calling system.
I think the thing I'm like very proud of what we did inside Block is that our agents can interact with pretty much everything we use as an enterprise to do work, so they're like extremely comprehensive, and I think we do a really good job of optimizing their efficiency and making them more and more ambitious in what they can do.
We talked earlier about we were really early on just thinking about agentic harnesses and like how that would be the way things would get done. I don't even know if you have an answer to this question, but why do you think we were so early to this? Is it like something cultural? Is it just like Brad like you and a few other folks like happened to be here at the right time? Why do you think we're so early to this like realization that we made a bet like a long time ago now and it actually was right?
Right. Yeah, I think there's like always a little bit of like, you know, serendipity there. I will say that this is part of the answer to this question is also why I really enjoy working at Block, and that's because I think as a company, a culture, like a leadership team, we are always encouraging everyone to be on the frontier of technology. That is not just like a kind of a notion or an ideology. I mean that it's specifically in the sense of if you are working on something that looks like it has potential, we will clear the time that you need to follow through on that. I think there was definitely like right idea, right time is what's part of it.
Another big part of it was that once we saw that there was some promise, we were able to carve out a small team and let them run with it, and this was not any of our jobs when we started working on it. It was like something that became an organization because the idea was working. I don't know that every company does that in the same way. Like, I think we have a long track record of doing this in different technology spaces at Block. It was this kind of expectation that as long as we were showing that there was potential, we'd be able to work on it and put the time it needed to like develop it into a real thing.
I think that also carried through not just in like getting that initial idea out into the world, but we kind of did this in, so one big part of this is that it's all open source. We, when we had this idea we got it out there and started sharing it with the community. At the same time as we open sourced it, we made it available to all employees, like not just engineers, and I think we did that really early and that has given us this like feedback system internally. Like 10,000 people have used Goose, internal, and that's something that we can iterate on.
Like again, I think that that's like, you know, that we bring that back to the open source version, but also like we were talking about inside Block it's connected to everything we do, and we got there because we encouraged everyone to use it and to tell us like what else could they do with it.
Yeah. You mentioned open source a few times. Like why did we decide to open source it?
Yeah. I think there's a couple of answers to this. One, as like a tech person myself, our whole company and the whole industry runs on open source technology. We get an incredible amount out of open source, and I think we're generally just excited to be able to contribute back to that community. I think in one way we were really motivated to open source Goose just to say, "We think we have something cool here. We want other people to try it and use it and like benefit from it if there's anything there." I think another real big angle to this is that we believe that by open sourcing a tool like this it will become a better tool.
Going into an open conversation, having contributors from outside our company makes it better and, in the long run, all of our best technology is maintained and developed by open source communities. I think it's just because we get so many ideas that, from outside like the scope of one company and that gives us this like long-term trajectory that I think beats proprietary tech most of the time.
Yeah. Yeah. All right. Maybe last question on Goose, which is maybe the most important, but how did Goose get its name?
Yeah. I feel I think I was I always hate saying the answer here because I think the cat's out of the bag at this point, so it is indeed a Top Gun reference. This was again like early 2024 we were talking about you know what would be a good co-pilot name? We were just-
Yeah
-like, in the office, like discussing, like, "Okay, what are some famous co-pilots?" We went with Goose. Since we did that though, I think the thing that really saved us on the branding is just that, like, Goose as the animal has, like, a lot of great, like content. I actually am glad we, like, leaned away from the piloting a jet and more towards Goose the bird.
Goose the bird.
Yeah.
You also, like, kinda look like Goose yourself, so I think there's a little bit of an element there too, you don't wanna sell yourself, but I think I could throw that out there also.
I actually did. I did do a Halloween costume from Top Gun and it worked pretty well, yeah.
Yeah, yeah. I can see that. Maybe, like, spending a minute on you personally, but like what's your role or what has your role been like in the AI industry? I think you were the first person to contribute to MCP outside of Anthropic. Like is that right? Just, like, talk about, like, what your role is in AI and open source, and getting away from, like, your day-to-day in Block. Like, I think you have-
Yeah
You and a few of the folks on the team have, like, a bigger presence just in the AI development space generally.
Yeah. Yeah, definitely. That's true. A little bit before MCP was publicly released, we were talking to Anthropic about agents and Goose, and saying like, "Hey, the main problem we're having with Goose is that people wanna do so many things with it. We need a protocol so that people can plug in new systems without us having to be a bottleneck on what they can do." We got to meet with the team developing MCP, so like David. They were like, "Oh, we have something for this. We're gonna release this." We, like, took a look, had some conversations about it, added a few contributions. Since then this is like a relationship that keeps growing.
We have some members of the Goose team at Block are on the MCP steering committee, and we contribute, and we kinda help build the Rust SDK, which is now mostly actually maintained by some fantastic contributors outside of Block. We got that started. More recently, the Linux Foundation has started a new sub-foundation, the Agentic AI Foundation. That is a home for currently MCP from Anthropic, AGENTS.md from OpenAI, and then Goose from Block. We were the three founding members there. I'm now on the technical steering committee. We'll, I think, in the near future be bringing in many more projects into that AI foundation.
I think the way to think about that foundation, by the way, is kind of helping to define what the effective stack is for building with AI. MCP is a really core component of that. I can talk more about MCP, by the way, if we wanna get into it, but MCP is a core component. AGENTS.md is a really, like a really successful specification for how people can customize how AI works.
Mm-hmm.
I think of Goose as the kind of reference implementation. Like it's an actual concrete thing that people can just go download and use that shows how all the pieces fit together. I hope that this foundation is the thing that over time does what some of the other open source foundations do, where it like kind of defines the and moves forward the stack that lets people solve problems with AI.
Yeah. It's worth, I think it's worth reiterating. I know you just said this, but three founding members, OpenAI, Anthropic, and us. It's pretty remarkable that there's, like, that level of, like in a relationship between Block and, like, some of the people that are driving the AI industry forward.
Yeah. I mean, it's an incredible opportunity. I think the role Block plays here that we do uniquely well is the application of the technology.
Yeah.
Like I mentioned, Goose is kind of the concrete of the founding projects. Goose is the most concrete. It's something someone can actually go use, and I think we do that in our products as well. Like, we've like Moneybot, Managerb ot, or Square AI, which has also been available for a long time now, are like really fantastic examples of some of the like state-of-the-art ways to apply AI to real world problems. I think that's a big part of why we're in that conversation is because we're helping to define like how to apply those technologies to become like useful products for our customers.
Yeah. Yeah. I think the element of, like, applying all this work to, like, actually benefit our customers is just so foundational to, obviously, to what we think about. I do think that there's something interesting about our culture here, where you talked about this earlier, but being on the frontier of technology, like just giving people the freedom and license that if there's something promising, going to do it and chase it down. Sometimes that will mean, like, things don't work, and you go back to the drawing board, or you go back to what you were doing before as like, "Okay, I chased this thing and kind of ran into a dead end." Then there's times where, like, a Cash App emerges, or a Goose emerges, or like-
Right
-there's really, like, foundational things to the business, and getting, like, the opportunity for people to work on that I think is just pretty unique here.
Yeah. Completely agree. I really do think that this is a thing we've internalized well. Like, I don't think we need to tell people to do this kind of work.
Yeah.
It's definitely not, you know, like. People just do it, and then sometimes incredible things come out of ideas that someone just had while they were working on their day job.
Yeah. I know we're. I'm just mindful of time. I wanna make sure like, a couple of things, a couple of other things I wanna make sure we talk about. We talked a lot about, like, I would say, like the compounding of the work that we've done in agentic infrastructure in the last, like two-
Mm-hmm
-or two-plus years. I would love to talk a little bit more about what happened in December and January timeframe around the foundation models themselves. When we've talked to a lot of folks about how we're thinking about the organization going forward on the investor side, it is a combination of like I think we're really early on the on agentic infrastructure and just compounding all that learning for the last several years. But then there's also this like moment in time where like just the world changed. So can you talk about like what happened in December, January? What's different today versus even six months ago? And like just talk about what that looks like or what that looked like.
This is funny because I think about this a lot, and if you've been in the space and like working with these models and you like look at the code that was written by the models before Opus and Codex 5.3 and then after, it doesn't look that different. You're like, I'm constantly asking like, what was it that flipped? I think what happened was that the models hit a level of like reliable execution that it kind of went from a thing that you had to watch and course correct, like similar capabilities, but a lot of failures.
Mm-hmm.
Now they have become a thing that more or less works in one go. That by itself is like useful, but it's not this like giant shift or this like phase change that we're experiencing alone. The reason why it became a giant change is because once everyone realized that they were like 90% likely to do what you wanted them to do, we've started to build systems on top of them that did more things in parallel or like did many executions in a row or ran it remotely or overnight. Because the reliability went up, that means that you can build these higher level tools that do way more work. We see that one really concrete example in something like Goose, but it's in most of the agent harnesses now, is this idea of sub-agents.
You have one parent that is like talking to the human about a high level goal, and then it dispatches to many individual sub-agents in parallel. Like we've had this idea for a long time, but it didn't work as well before because they would fail more often than not, and it was just kind of frustrating and you had to like micromanage them, right? Now you can have one agent orchestrate 10 agents, and they mostly do what they need to do. You've got this like kind of 10X multiplication there of what it's gonna take on through this like higher level system development.
I think that's the phase shift that is really happening, and you see that in all of these like new tools that are trying to help you manage running like 10 agents at a time or, a lot of people posting on X about how they've got like 18 terminal windows open that are all working. That's what changed, is that they're all so reliable that you can be running a lot of them at once, and we're just like scaling out how much we can run agents to do. For me, that means that like you were saying with like something like builderbot, we can have someone just ask for a feature on Slack, and then it just goes and works for an hour and comes back to you.
A lot of the time, it did what you wanted it to do, and it's like high quality work that you can just go start merging into production.
Yeah, the parent and like sub-agent feature I find fascinating. Like, what I was running an analysis earlier this week, and I just asked Claude like, "What are you actually doing right now?
Yeah.
Something like that. It said, I have it right here. It was like, it said, "I'm in the spotting phase. I have four analyst agents launched in parallel.
Yeah.
Segment analyst, a risk impact analyst, a benchmark analyst, and an operational analyst." They were all doing like different components of the research I wanted it to do.
Yeah.
I came back and it was like a really good research report that would've taken me like, you know, I don't know, hours or days to do, and it was just done.
Yeah.
Um-
I think the real thing that's happening is they used to try that idea as well, but it was like chaotic.
Yeah
Because those sub-agents would fail so much that it would just kind of devolve and spiral. I think the real flip of the switch was that they started succeeding more than failing, so that instead of it being chaotic, it's mostly self-correcting. Suddenly you can just do these incredibly ambitious things.
Yeah. Talk a little bit about your day-to-day. I wanna talk about product development generally, and I, and we'll probably bring in like builderbot and those kinds of things, but maybe just your day-to-day. If we go back to, let's say like December 2023, maybe like right before Goose, like before you started maybe spending all, most or all of your time on building Goose, like what's your day-to-day like today versus then in terms of just like building stuff and how the world's different for like a development organization?
I think It's like it's so incredibly different. I think most people have this experience now. An example of this is that I have literally uninstalled all of my code editors. Like I do not write any code by hand anymore, and I think that's really common. And so that like it couldn't be more night and day, right? Also another thing that is an example of that is that I no longer care what programming language I'm using, which is just so enabling for me as an individual because I like my background is mostly in Python. I still think I can write decent Python if I need to, but I've never been any good at front-end development or TypeScript.
That's something where I can just go build interfaces now, and it doesn't matter that I'm good at TypeScript. It matters that I know what I want it to do and that in terms of like writing high quality production software, it also matters that I know how to test and like how to confirm failure modes and like make sure things are working the way that they are expected to work. Those are transferable skills from one language to another. That's, I think, like, a huge difference is that I write now in like five or six languages a day, and it doesn't. I don't need to be an expert in any of them individually. There's so many more things like that in software development that have changed.
The other thing is that I'm multitasking like crazy. I think this is true for everyone. That's not just a developer thing now. Like, I will have, I always try to have, like, seven or eight things that I could work on at any given time, and when I have a new idea, I just go find one of the interfaces that I'm using to manage agents and then go send one off and let it do the work. I come back to it an hour later and review what it did. It's a lot more of this, like, kind of like working in an organization in some ways.
Like, where you're talking to multiple people, they're all doing work, and you've kind of moved up one level, and it's almost like you're orchestrating a system of agents rather than looking at, like, individual code changes. Huge difference in software development. That's the question of, like, the how. The what has also changed a lot. Obviously, like, we're as an organization, applying AI into our internal productivity and our products is, like, one of the most important things we do now. A lot of the what has been shifted towards solving problems in our products with AI as a foundational building block.
Like, if you think about new feature development, you like, there's a lot of things that weren't possible before that are possible now because AI can, like, bridge from what the human wants into what our systems need to do.
Yeah. That's probably a good segue into, like, the next area I wanted to talk to you about. I think we talk a lot, like, we've talked a lot externally and, like, just conceptually, like, velocity makes sense, where it's like, I have an idea. I'll spend like 20 minutes, like, writing it on paper, and then I just, like, have someone, like, have this agent build it, and it's, like, a pretty good prototype that I can then share with somebody and use, like, a conversation about it. Can you talk a little bit about, like, the quality? 'Cause I think that might be probably-
Mm-hmm
maybe it's as important, but probably less well understood.
Yeah.
there's the prototyping velocity and so on, but as you get into, like, actually shipping things into production and just, like, the quality of what you're able to build, like, how does AI impact that or factor into that?
Yeah. I think this is maybe the next frontier in a lot of ways. Like, prototypes are getting close to free. You can just do this incredible product life cycle where you're always communicating in prototypes. You see exactly what it's gonna be like for the customer. That has changed how we work. Now the question is, once you have the prototype where you think it nails it, how do you go make that quality that you can actually deploy for customers and that it's gonna work reliably? Some of that is, like, there's a mix. In some ways that's actually the new bottleneck, right?
Uh-huh.
Which is just, like, people are reading code still to make sure that it works. That's usually the expectation. There's a lot of higher level thinking too that goes into that, and testing. Like, how do you build a test set that proves that this is gonna work reliably? There's a lot of, like, experience and intuition into building that test set. That is why I think engineers are still, like, a critical part of how we do the work. Even if code is, like, easier to write, it's coming up with the systems that make sure that it runs reliably that is a big part of the job now. There's part of that. I do think, though, that agents are also helping in that part of the process.
We have, like, code review agents that help us, like, find bugs that maybe a human wasn't paying close enough attention to see. That's hugely additive because it just again, it's just, like, there's, it's all bonus, right? Like, if you catch a bug with an agent that a human missed, that's just one problem you didn't have to fix later.
Right.
There's a lot of that. We're looking at also, like, techniques where the agent helps explain to the human what the code does, just so that their life cycle of reading the changes is faster. It's like, "Oh, here's an intro to what you're looking at," that speeds up that initial process of getting up to speed. I still think currently, with the current models, I don't really trust them to be the designers of the tests of, like, how do we know that this is working? I find that they tend to make tests that miss the point still. That's a huge part of, like, what the engineering teams are doing, is like, what is the thing that proves that this works and scales and will be reliable for the next year?
Yeah. All right, so maybe rounding into, like, the last general area that I was hoping to talk to you about. If I think about, you have the foundation models as, like, the base layer. You have the agent substrate, which for us is Goose, and everything that we've been talking about there. Then there's the end, like, applications of getting a lot of this work done, or, like, more as, you know, more importantly, like, getting it out to customers, like, making their lives better. Maybe on the getting the work done front to start, can you talk a little bit about builderb ot? 'Cause, you know, like, I had my first experience with it today. It was really cool. What is builderb ot? I think you're spending a lot of time on that now.
Like, what is that tool, and how does that move the ball forward for Block?
The three bots. I love the naming on this, by the way. I think this was a Jack invention. You know, Moneybot is for our consumer-facing automation. Managerb ot is for business-facing, like Square sellers. Then builderb ot is for us internally. Under the hood, like we've been talking about Goose as the agent harness, and Opus 4.6 or GPT-5.4 or GLM 5, those are examples of the LLMs builderb ot uses those two things and then probably the easiest way to think about it is, like, the collection of capabilities that you bring into the harness. Like I was mentioning before, a lot of the design Goose can run any kind of tool.
A lot of the work for internal automation is figuring out what tools it should have so that you can actually do the things that Block employees need to do and that's what builderb ot is, it's those, like, kind of, all the glue and bringing it together all the capabilities that we need and then making it convenient to use. An example is we run this on Slack, right? That's where most people get first introduced to it. Incredibly powerful to be part of that early conversation, like when people are working with each other to just be like, "Hey, the bot, how does this work?" Or like, "What do I need to know about this?" That's already had this enormous impact. That same thing also is, like, becoming very rapidly successful at reducing operational load.
An example of that is, like, say that we're having someone is coming to ask for help from a platform team or something like that. Like, builderb ot can be the first thing to answer because it knows all of the code, knows how everything works and so it's just kind of smoothing out those, like, operational burden. Also true for incident management, for example. Like, one of the main ways we make a reliable system is you don't really plan for things to always work all the time. That's unrealistic. Instead, what you optimize for is how quickly you can fix any problems and so now what we have in most of our systems is that builderb ot, as soon as an error is detected, builderb ot will, like, wake up and start triaging.
By the time a human manages to come find, like, and connect and look at it, hopefully there's already an analysis of what's wrong and how to fix it, and that's actually been working incredibly well. Yeah, that's the way to think about builderb ot is it's an instance of Goose running on a frontier, very powerful model that has all the tools it needs to do work at Block. There's a ton more to this that's really cool. Like, I'm very excited about our new remote workstations, so if you want.
Like I was mentioning, I now work on a lot of projects all at once, and so my laptop becomes, like, this limiting factor on how many things I can do at a time 'cause it's like, you know, you can only compile so much at once. We now have these remote machines that builderb ot will go run on and, like, do work for you. You can do that overnight, like close your laptop and come back to things in the morning. It's kind of that a lot of new interfaces so that people can have agents running more automatically or at higher scale, and then all the tools it needs to do the work.
I think I need to talk to you about those remote workstations 'cause I was building the thing with builderb ot and I was like, "You should use that, but you don't have access to it yet, so we can build it locally and it'll be just a little slower." I'll talk-
Yeah.
I'll bug you later about how to get in there. We've talked, okay, everything we've talked about is like kind of been internally focused, I'd say, but you've mentioned Moneybot and Managerb ot a few times.
Yeah.
I think ultimately the most valuable thing we can do for with all of this technology is say, like, you know, deliver it to our customers and help them manage their businesses better or manage their financial lives better. When you think about, well, maybe like six years back, just talk about what Managerb ot and Moneybot are. I know we've talked a little bit about that. What roles do you envision those playing for our customers, you know, 12-18 months down the line?
Yeah, right. I think so. Okay, Moneybot. It's in Cash App. It's rolled out fairly widely now. I think it's-
It's about 20%.
Maybe not everyone. What was that?
I think it's 20% rolled out now.
Yeah. That's right. Yeah. Not everyone has it, but a lot of people do, and in Cash App, you can click on it. It's got a nice little smiley face, and you click on it, and it has access to basically all of the things that you can do in Cash App. Like, more or less, Moneybot can do everything in Cash App that you can do. You can ask it questions about, like, how you're spending money, how would you go about setting up, like, a savings goal. It will actually, like, help you operate the interface. Like, if you can't find the right place, it will navigate you to the right place, and I think that's, like, the foundation.
What we've done really well right now is this like foundational setup where Moneybot can control basically anything you can do in Cash App, and we've done, I think, a really incredible work on making that intuitive and a system where it never does anything you wouldn't want it to do, right? Like, it never moves money for you. It instead shows you a really clean interface where it will tee up a transaction, but you as a human are always involved before you know anything happens. You approve it or you see this like nice confirmation page. That's some like I think we're doing really cool things in the generative UI space around that.
Let me cover Managerb ot, and then I'll talk about where I think they're both going. Managerb ot is the same idea but for sellers. With business operations, I think it's a slightly different space, right? Because when you're working on your personal finances, it's less data and you want it to be faster. You, it's really nice to have this, like, fluid conversation, really get insights quickly, take actions quickly. Like, if you just wanna deposit $200, you want that to take, like, a couple of seconds. We do a really good job of making that fast and efficient. With Managerb ot, there's a lot more of the like, "Oh," like, this is something that might take a couple of hours.
We're also looking at interfaces there that push into that frontier in the same way we do for employees, where it's like, "Oh," like, "let's let that run remotely and take a while and manage these, like, long-lived tasks or do them proactively for you before you need to do them." Because, you know, when you're running a business, there's just a lot of work to do, and all of the things that are on the back end, we would rather have a bot able to do that for you so you can focus on, like, the parts that really matter. The both are the same ideas. It's like AI that can operate our systems and make it easier for our customers to do, to focus on what they need to focus on. Okay, where are those going?
I think there's two angles that I'm really excited about in both of these. One of them is the proactive intelligence layer. What that means is that, like, it's going to take actions for you before you need to ask it. Like what we find a lot of the times is you give people this empty text box, and they don't know what to do with it. How can we actually just like, because we have this existing relationship, how can we suggest to you what would be useful for us to do?
Even increasingly, like prevent bad outcomes for you, like stuff that like, you know, maybe if you're a Cash App customer, and you have an upcoming bill, could we like prompt you to deposit some money so that you don't, so that you've got enough in advance for that like Netflix subscription, that kind of thing. Managerb ot, same thing, but more complexity. Like how could we just like help you with the operations, like manage upcoming appointments or booking your time cards, that kind of thing. That proactive layer, really cool. I think there's a lot we can do there, and we're already starting to see the value.
The next thing is like giving you this generative experience and there's a lot of intermediate milestones that I think will be really cool, but the long-term vision of that is that you as a customer should just be able to ask us to do something for you, and we will on the fly use AI to make it happen. You don't have to wait for us. You know, you don't have to send us feedback and ask for a feature. You can just come to these like agentic interfaces and say like, "Hey, I need something that does this," and we'll just build it.
Yeah. That to me sounds somewhat like science fiction where it's like you're a Square seller, you have this like really nuanced, like inventory piece of software that you need for, you know, maybe your business is like really specific. The idea that rather than waiting for Square to build it, you just say like, "These are the specifications I need, this is my use case," and it actually just builds it for you. Again, it feels science fiction-y, but it doesn't actually seem like it's that far into the future.
Yeah. I honestly think that even if we never get a better model, we can make that happen with the current models.
Yeah.
It's about the kind of agentic engineering that we need to do, but I'm confident we can do it.
Yeah. Well, Brad, we've covered a lot of ground. Anything-
Sure
Anything you think we should have talked about that we didn't?
No, this is great. I think maybe just the one thing that I would say more about proactive intelligence is that I think this is where a lot of value is because it's really like there's a lot of. I really think that the we have a history of doing this well, and that it's probably one of the things that we can be most competitive as Block. Just to give you an example, like we are, I think, fantastic at risk, and that is a form of proactive intelligence, right? Where if you're making transactions, we protect you from losing money by preventing fraud on the platform. It's a kind of proactive intelligence that when it's working well, it's completely invisible.
Yeah.
I think we're excellent at it. Lending is the same way, right? Like the fact that you don't have to like reach out to us for a loan, we can actually just like offer a loan proactively based on your cash flow. That's an example of proactive intelligence that we've been doing for years. I think the thing that we're that's like a thing we're very good at as a company. The next step to me is just connecting those same ideas into the like much more powerful surface that AI can then go take actions for you.
That's something that I think, again, like, you know, we have this existing relationship with so many customers, and I think that layer is how we're going to take it and turn it into these experiences that you can't do without that background relationship.
It does feel like we've talked a lot about proactive intelligence, you know, like Jack and, you know, externally and internally has talked a lot about it. Do you feel like that's like is that focus unique, or is just like our capabilities and the fact that we've kind of been doing this for years and years and years with or without AI, and we can now like do it with an AI, like does this make this is our ability to execute on that vision, do you think, relatively unique among the peer set that you observed?
I think it is unique, and there's a few reasons why. One is the like fundamental data that we have as a company that other companies don't have. That's why I think a startup would have a really hard time even approaching this problem today. That's necessary but not sufficient. The other thing that we have to be good at that I think we're proving we are good at is being on the frontier of applying AI to problems and figuring out how to like so turn all of that data that we have into like real world solutions with AI.
I think we're showing, you know, both ends of that very successfully today, and I think that gives us the best opportunity to be the first company that figures out how to bridge them together.
Yeah. I think that's a great way to end it. Brad, thank you so much for the time. This was awesome. Really enjoyed the conversation, and I'll talk to you soon.
Yeah. Thank you so much, Matt.