Angie Jones on Ralphing 25k repos at Block, GPT-5.2 Codex, and CES weirdness

0:05Welcome to another Friday edition of Dev Interrupted. Our second one, in fact. I'm your host, Andrew Ziegler. And I'm your host, Ben Lloyd Pearson. And joining us this week is Angie Jones, the VP of Engineering, AI Tools, and Enablement at Block. And Angie, I'll be honest, we've been dying to get you on the show ever since you and I collaborated on that CoTV episode together, where we worked with Goose to create an app that you didn't interact with in a normal way. It was a ton of fun. Dropped back in October. And can you believe that that video is at like half a million views now? It seems like everyone is goosing these days.

0:43It was a good time, for sure. It was a good time. So, you know, this news segment, I thought this would be the perfect time for you to come and join us. So thank you again, Angie, for being here today. Yes, of course. Thanks for having me. Amazing. So I want to dive right into our news segment and our dedicated episode now. So a big theme in the past couple weeks at the top of the year has been how developers have come back from their winter break in a whole new mode of agentic development and how they're taking advantage of it. And we actually recently had Jeffrey Huntley, the so-called inventor of the Ralph Wiggum loop, on the show last week to talk about where it came from.

1:20We'd actually been covering Jeffrey for the last year on the show and the things he'd been writing about. So when Ralph became more mainstream, it was really exciting to get together with him and see where his head was at. You know, in Jeffrey's work, it also is incorporated partly with Steve Yege's work on Gastown, basically taking a whole army of Ralphs to do things for you. And we're discovering all new modalities of engineering there. So, Angie, I know you've been following all of this, too. I've been following the things that Block has been posting. And Block, I think, has been really meeting the moment with Goose and making it something that you can Ralph with.

1:53and explaining that Ralph is like a technique that carries across so many things. And it's not something that's unique or special. It's just clawed code. So, you know, tell us a bit about AI-assisted development at Block and what y 'all are focusing on right now. Oh, yeah. So we're pretty mature, I would say, in our AI-assisted journey at Block. We have about 95 % of our engineers that are actively using these AI tools all the time. They use a little bit of everything. So there's Goose, there's Cloud Code, there's all sorts of other tools. Some work better for, you know, mobile, you know, because we're a big shop where we have front end teams, back end teams, you know, iOS and Android.

2:38And so people get to pick their flavor. And yeah, they're using these tools regularly to do amazing things. When I saw the Gastown article, I couldn't help but say, oh, let me let me look at the numbers and see where my folks fall in all of this. And I was really happy. Now, when you say the numbers, you say the charts, the graph of like one. No, so I actually, you know, because I'm responsible for AI enablement, I actually can pull the data, you know, from our own engineers to see how they're using AI, right? Which tools are they using? Do you know, do they are they in IDE? Are they using CLI?

3:21Like what's going on? Right. And so that gave me a really good picture of where we were on that on that Gastown graphic of stages one through eight. And yeah, most of our folks are about at five. So it's happy we weren't in the top row. We were about at five and six. And I'll tell you, I have, I'll take a step back. I'll come back to that later. But most of the folks are five and six. And then we do have like a very small set that are at like the, we're building our own orchestrator level. Or they're using something like beads and things like that, right? So a lot of what got us here, one, I'm going to give credit to the models.

4:04The models just got a whole lot better. I don't want to take all the credit. But then, but also, I started an AI champions program. And this is 50 engineers from across our business. Each one represents a different repo. And we think repo readiness is a key to a lot of this and understanding context, engineering techniques, and things like that. And so that group, they really went in. I got buy in from their leads to say, hey, I need 30 percent of their time to really invest in getting AI and integrated at the repo level. So it's not just individual people benefiting, but an entire team. And that got us really far.

4:45So those people are the ones that are like kind of the six, seven stage camp in the in the gas town. And actually, I want to call that out because we read this article that you published that we'll share in the show notes. And there was a line that stuck out in that article that described these people. And you said they were resilient engineers who, when AI hallucinated or produced slop, they leaned in and figured out why it failed and work on the engineering to make that a success. And I think that is such a critical approach to this because a lot of people had kind of the opposite reaction. They saw it fail and they thought, well, this doesn't work, so I shouldn't use it.

5:21And it's really clear from reading this article that y 'all were very early to the AI enablement function because you were doing practices that I remember a year, maybe as far back as a year and a half ago, we were also starting to establish. And it was just to deal with a lot of the realities of today's tooling. But yeah, and there's a lot of practices you outlined that are amazing too. that I really appreciate too, like the RPI research plan and then implement. I realized that I have taken that approach to a lot of problems as well without really having a term for it. So yeah, I just wanted to call that out.

5:54The Champions Program was a really great idea. Yeah, and shout out to HumanLayer for introducing that RPI pattern. We found much success with that. We're getting really high quality PRs. There's also something really unique I want to call out about Block that makes it a little different and how technology is adopted there versus other companies is like you called out in the beginning block is a shop with a lot of different tools a lot of different ecosystems of technology different high levels and low levels of programming there you have a large open source like library that you maintain that has all sorts of different needs so your developers are not only at all these different skill levels and adoption levels but they also are on this big gradient of ai does work for their specific industry or their niche or something that they're focused on.

6:40But for others, it can be really detrimental. And meeting that challenge as an engineer and fixing away those failure modes is, you know, it's hard. So you kind of have like a really broad view versus like companies where they maybe are lean very SaaS forward, very like more modern languages, they might be able to go faster? Yeah, that's true. We have iOS teams that really struggle with a lot of these tools, right? They work very well for front-end team. Like a front-end team might say, oh, we cracked the code. This is what you need to do. This is the tool. This is the pattern. It works. We're getting 95 % success rate, right?

7:20And iOS goes, oh, wonderful. And they try it and they get like 20 % success rate. So, you know, there's variance here. And what this champions program allowed us to do is kind of stress test this across different realities to see what actually holds across the board. And when you find what holds across the board, you get these emerging primitives, like these basics of working with the tool that everyone should pick up and apply. And I think that's where Block is doing a really great job of finding those, digging those out from your own internal audiences and from the external ones, too, and sharing them.

7:52and it makes me wonder too like in this world now what we've been talking about like ai works at older sorts of different levels you got to go in as the engineer and solve the failure modes enter ralph enter the ralph loop and then that kind of like puts all of this on a hyper speed and all of a sudden you have all sorts of people trying to apply this at different levels of success but what you are seeing more broadly across the board are people finding the language that empowers them to explain that you throw down the hallucinations and you fix away the failure mode. So what do you think of like what Ralph is doing right now to how engineers are approaching their work?

8:27You know, it's funny because it feels like a meme on first sight and you go, ha ha ha. But again, our folks were like, we were, we're willing to actually like look through and try things, right? And go, okay, a bash loop. Like that's not, that's not silly, you know? And so we do use it in different contexts where it makes sense. Where does it make sense to maybe have a bash loop. For example, just this week, we have like 25 ,000 repos, right? And so one is like, one-to-one initiative is, okay, everybody, we need to get your repos like AI ready, right? And so that's a tall task to then go to every repo and like, all right, we got to like figure out what rule files you need, da-da-da-da-da.

9:16So we actually are doing an initiative this week where we brought out Ralph. And we said, hey, Ralph, look, go to every repo, assess the repo, and then build out an AgentMD file that's specific to that repo and go ahead and put up a PR for it. So beautiful use of Ralph, you know? Yeah, just set it on itself, right? It's a very solvable, deterministic, very almost like trivial problem that you just need like a large amount of time for. It's the perfect thing to throw a loop at. That's an amazing way to meet that challenge. So obviously, there's lots of internal brainstorming about what it means.

9:55And then when you take the idea of like Gastown, for example, which evolves it, like you now you have like a village of Ralphs or however you want to, you know, characterize it. I definitely think the writing is very flavorful in terms of its metaphors and making up lots of things. But I do think along the way, there's things that emerge that are very useful. There's one thing that you mentioned that I want to shout out to an early kind of emerging thing we're learning from the realities of Gastown are things like Beads and why a library like Beads is really helpful with working with AI. I actually have adopted Beads in the last several weeks and I have been using it daily and I am a huge fan of the Rust port of it by Jeffrey Emanuel as well.

10:36I use that in my workflow and what I think it does actually is it allows the it's like a thought substrate between you and the LLM. So this is like it's almost like a protocol level discovery I think between what we're doing in Gastown that only comes out of working at that speed and at that level. And it solves a lot of interesting problems. So I think that like the faster we go and the more we accelerate to the more interesting solutions are going to crop up just from the realities of it. That's why when people are like, oh, like software development is dead. I find that ridiculous. Like I'm seeing new challenges every day.

11:13Like as soon as we figure one thing out, that opens up, you know, a bunch of more challenges. For example, like now we're cranking out PRs like on this automated way. Who's going to review them now? And, you know, we work in finances. This is not a just ship it kind of thing. Like a human has to review this stuff, right? And so now that's a new problem that we're trying to figure out how to do this in a smart engineering way, right? Yeah, and actually this kind of gets to something that I'm curious about because when we think about like Steve Yeggie's eight layers of agentic development, right?

11:47And it feels like once you've achieved that eighth layer, like if you put in all the work to sort of build your own gas town and have this capability to like essentially direct an agentic army at problems, the next challenge in my mind seems to be how do you get multiple people operating at that level to collaborate together in an efficient way? And then how do you get them to communicate with each other? I'm always talking about the handoff to handoff of the situation. I'm like, what, is my gas town going to call your gas town? Like, I don't think that's the reality that we're heading towards.

12:21It really calls a lot into question about like how teens operate. Yeah. Yeah. So I'm wondering if that's something that you've like sort of started to see at Block. Yeah. I mean, it's something that I have to keep at the forefront of my mind because while all of this stuff is exciting to read like on the weekends or in the evenings, my reality is that I need to get a whole engineering org at that level. And this is not a one by one type of situation, right? This is very much a team sport, which is why most of the time I'm targeting things at the repo level, because that automatically kind of just unlocks and gives the entire team these capabilities, right?

13:00And then, yeah, when you get into like the orchestra thing, okay, like, so now I have like my multiple things of agents, we're actually working on something for us internally, where we want this sort of, I can see, I can see your, your task that your agent is doing, you can see mine, they can collaborate together, maybe, you know, they pick up a piece and, and your, your tell your bot to call my bot sort of thing, you know, and they're all working together in this collaborative, like orchestration space. This is definitely the year of interfaces, it seems like, already, where everybody is realizing the tools we have now don't really support the way that we've come to work.

13:39And so I'm excited to see what folks come up with. I completely agree. I've built three new tools already this year. It's the third week of January, and I'm on my third or fourth UI of the year. You're literally not joking. He checks his watch to see when his clawed token, or his clawed I got my burn. My burn rate is everywhere on my computer. I know what I'm using. I love it. But I think this is a great way to segue to the last thing that I sort of wanted to talk to you about and how this this impacts the overall software engineering industry. You know, so we have a lot of people out there that I mean, this this keeps coming up time and time again.

14:17Like we just saw this. We'll leave a link in the show notes of this message on X from Ryan Dahl about how, you know, the era of humans writing code is over. And, you know, I think that is, in a sense, it is a reality, at least the way we used to have humans writing code is certainly many aspects of that are ending. But I'm wondering, you know, as somebody who sees large scale engineering organizations operate, like, I'm just curious to see what your perspective is on, you know, I think a lot of people agree, like software development as a profession is probably dead. But what does software engineering look like in the future?

14:54Yeah, I think I think the coding pieces is probably accurate. Right. Just for funsies. This weekend I coded something or last weekend I coded something by hand. You know, I'm like, all right. Yeah, I still got it. But this was such a waste of my time. You know, it wasn't even as enjoyable as it once was, because now I'm like doing the grunt work to get to the solution. Right. So that part, I think, yeah, we're about done with that. We're at a higher level of abstraction. That doesn't mean engineering goes away, right? You still have the engineers kind of basically surfacing what it is that they want done and maybe how they want it done and things like that.

15:37It's just you don't type the keys anymore to make that happen. And they were going to unlock just all sorts of new challenges. I mean, that's the reality of it. Like, how do I work with 10 agents in parallel simultaneously to knock out what would have taken two weeks to do in a day? Right. That's your new engineering challenge. Doesn't that sound exciting? Yeah, exactly. How much time can you save yourself and other people? How much time can you give yourself back? That's like the new challenge. Yeah. And I think it's been the easiest way that I've been able to get people on board to this AI hype train is just to show the amount of toil that we can remove from their life.

16:17Like once you start to understand it is that grunt work that is like the first thing to go away in this new model, then yeah. And it frees you up to start thinking about those higher order challenges. so amazing well there's a lot of like places where this can go but angie i think ben and i could probably talk to you all day so i think we're certainly going to have to have you back in the future to go deeper on some of these topics especially learning more about how block is meeting the moment and all of these different challenge domains of adopting ai and of course building goose an amazing operator an amazing harness that i use all the time i'm actually using Goose later today with Rizal on a live stream with her.

16:57So I'm a big fan of it. I definitely recommend folks go check it out. But Andy, before you head out, I just wanted to ask where people can go learn more about you and any final thoughts that you want to leave us on. Yeah, thanks for having me. Yeah, you all can go to my website, which is angiejones.tech. And all my social handles are there. I'm doing a lot of blogging these days about these things that I'm discovering and learning. And I think that's the key. We are all figuring this out as we go. So I love when people are sharing their tips and tricks and I'm going to share mine as well. Amazing.

17:33Well, thank you so much for being here, Angie. And we'll have you on the show again very soon. Take care. All right. Take care, guys. Well, we have a few more news stories to roll through, but it was amazing having Angie here, Ben. What did you think about some of the things that she said? Yeah, I mean, it's incredible to get the perspective of somebody who is seeing the changes that we are witnessing at the individual level, like hitting on like thousands of developers. And, you know, and one thing I just want to reflect on is kind of what we closed on about how this is impacting the larger software development industry.

18:05And, you know, I've started to sense this like growing tension between engineers who have really started to like realize the productivity benefits of agentic AI software development. But there's also a lot of engineers out there who identify themselves with their ability to craft finely constructed code and solve interesting and challenging problems. They spent years, potentially decades, building up that capability, and that's sort of what their identity has become. And I think there's a lot of doom and gloom with that second group, but I don't think it has to all be doom and gloom. And, you know, as we talked with Angie, like, you know, a lot of those jobs that are just low level turning tasks into code, like that type of job can be done by AI now.

18:51And if you are in a job that is due, that's your responsibility, it could become irrelevant in the short term. And I think that's the dark side. But I don't know if there's actually a huge number of developers out there who work that way anymore. You know, the vast majority of software engineers, you know, have higher order challenges that they have to solve, too. And I actually think there is a bright side to this. Like, I think there is still a lot of opportunity for the people who like to solve extremely difficult challenges. You know, LLMs are really great at replicating, like, past success, but they're not so great at innovating new things.

19:25And I think there's still a lot of room for people to be at the cutting edge and focus on writing code that creates innovation and solves difficult challenges. and there actually may be even more demand for those capabilities in the future as this technology develops. So, you know, I think there's a lot to reflect on our conversation with Angie and yeah, I'm going to be thinking about this for a while. We definitely have to have her back. For sure. I will be thinking about it too. And if you are a listener, you know, you have some thoughts, please be sure to drop them on our LinkedIn or wherever you can find us because we'd love to know what you thought about that as well.

19:57But I want to dive into something for our next segment covering some news from OpenAI. And this is OpenAI releasing its GPT 5.2 codex and its response is API. So this means that you can call this flagship coding model, you know, through their API and use it in a variety of ways. This allows you to build with it like a building block and use codex to, you know, extract great value from how you work with your computer. So, you know, OpenAI, this is their most advanced agentic coding model. I think it's really important to flag here, you know, something Angie said about like recently the models got really good.

20:33I think we're seeing that here. Like Opus came out in like, what, November of last year? Now Codex very similarly, and now it's available. It's like they're neck and neck in terms of their capabilities. And it's fascinating to see what people are doing, you know, with the Codex model. But I think that all of the competition is quite good. Ben, what do you think of this news? I echo your sentiments there. The competition here is really amazing to see. And I think it's what is making this moment so exciting for people who are leveraging all these tools. You know, I use multiple frontier models because I've found that they're more capable at certain things versus other tasks.

21:14And I'm actually curious, Angie, like what models are you using today and like what purposes do you define them for? oh that's a good question i use um i use flagship models from a lot of the different of like primary companies for different tasks so for example i do often go to gemini for things like image creation because obviously nano banana is just top of its game right now and i love using that thing it's a big life hack that not not everyone has learned yet that like don't don't mess with anything else for image generation go straight to just go go straight to general So there's that, obviously.

21:49I will say I probably do the majority of my coding with a Claude model. I love using the new Opus model. I think it's amazing for writing code. But that said, the GPT 5.2 model and GPT 5.2 codecs as well are in my tool belt. I use GPT 5.2 and ChatGPT all the time to do a lot of my copywriting work, to do a lot of cleanup on information that I put into it. And I actually probably use it as my primary daily driver for things that are not coding. So, you know, they all kind of have a different place in my heart. And I swap between them all the time because what I learned is that it's about just discovering the challenges that come with working in them in different ways.

22:30Yeah. And many workflows can include multiple models, too. So, you know, you can you just learn how to get them to interact with each other. But I like that you called ShadDB to your daily driver because that's exactly how I view it. It's like that reliable car that just gets you where you want pretty consistently and always sort of like it just works the way you expect it to work. So it probably does get the most usage day to day for me. I really like Claude for – I view it as like my builder. It's like when I need to like – when I already know exactly what I want because I've spent the time building that context or whatever.

23:06And then I just need to create something whether it's code or content. content clod is just really good at just bringing you from like a fully fleshed out idea to something that is like ready to release to the to the world uh and then yeah gemini aside from image generation i've also found it useful for i give it the big hairy problems like stuff that just takes a lot of digging to like fully solve you have a lot of data and you need to come all the way through it to find a specific problem with with every single occurrence uh gemini does a pretty good job it seems like at doing that. It's a little wild sometimes.

23:44It goes off the rails. I feel like a lot more frequently than the other models, but when you're solving big problems, it seems to be the best. So yeah, I think the competition is great. I hope it keeps happening for a while because we're the ones that are reaping all the benefits from it. And if you're listening to this and you yourself are not somebody who's experimenting with a lot of different flagship models on a regular basis, this is your call to action to do so. It's really important to rotate through the different offerings that are provided and understand what they're good and bad at, because that's going to increase the surface area for what you can do with this kind of technology.

24:18So really curious to know what your daily driver is as well. So be sure to let us know. All right, let's close out with a fun one. You know, CES just wrapped up in Las Vegas. I feel like there's always a lot of weird and strange things coming out of CES, much of it fake. But we've got an article we'll share in the show notes about all the the weirdest tech of ces and i'm curious andrew uh what was your favorite uh item from this article oh yeah i love this roundup you see yes you're so right is is very bizarre it's almost like vaporware but for like real life products you go and see things that are never going to come to be there are a few things there that stood out to me but the one that was the most straight out of science fiction for me uh were the lollipops that played music through your teeth into your ear while you ate them.

25:06That blew my mind. What I think we saw is huge leaps in the kinds of products that were getting put out, maybe because a lot of these production environments, a lot of these real-world product creation processes have been implementing AI and experimenting and trying new things. And so you're getting a lot of creative outputs. What stood out to you in this roundup, Ben? Yeah, I've heard multiple people tell me they want to go buy those lollipops now. I can't really get them out of my head. I feel like almost they would hurt. Like I have to try one. Yeah, this actually wasn't too weird to me, but this article link had, it mentions a product I've actually been following for a while because I've wanted this, but it's a hypersonic knife.