Ralph Wiggum goes to Gas Town and the death of the IC

0:05Welcome to Dev Interrupted, sponsored by Linear B. I'm your host, Andrew Ziegler. And I'm your host, Ben Lloyd Pearson. And welcome to our first ever Friday edition of Dev Interrupted. And Ben, I was thinking, should we name these Friday episodes something different? Maybe give them their own brand, make them something catchy? Like, okay, here's some ideas. I know, right? So there are some ideas that are already here on the script that I want to call out. So we got the Friday deployment or the hot fix. I like the Friday deploy, though. That's definitely a shout-out to some of the stuff we've been covering recently about continual merge.

0:45But what about DevInterrupted 2 Interrupted? Or 2 Interrupted 2 Furious? Wait, I don't know. I'm maybe seeing a trend here. What do you think, Ben? We're not doing Fast and Furious references, I think. What about The Interruption? This is your Friday interruption on DevInterrupted. Oh, The Interruption. You know, maybe some of these will stick. If you're listening, if you have a favorite one, please shout out and let us know. Yeah, yeah, yeah. If anyone's got great ideas, because naming stuff can be hard. But Andrew, you know, what can our listeners expect by joining this episode? You know, this is the first one we have.

1:20We're going to be doing this moving forward. So what can people expect from these episodes? That's a great question. So, you know, we get a lot of interest from industry experts who want to come on the show. We're always looking for ways to share more voices. And by expanding our new segment and giving it its own home on Friday, we can make that great space possible to bring in more expertise and share them with our audience connect you with the things that matter because really it's a sign of how much Devontraptit is growing and that's because of you our listeners being a part of our journey and supporting us and you know we piloted this idea over the summer Ben actually while you were out on a parental leave I brought in a whole bunch of guests news hosts and we cover all sorts of things from across the industry.

2:00And the energy was amazing. I want to channel that all year long in 2026 and beyond. And, you know, with how fast everything is moving in AI, it'd be really exciting to get people in here more regularly and share ideas faster to talk about what's happening week to week. Yeah, you know, I really loved hearing some fresh voices coming into the show while I was out. So that was a really cool experiment to run. And I think probably some of the people would try to get back on the show, you know. But, you know, I think really what we want to get out of this is that we are getting so much interesting things coming inbound to us, you know, whether it's white papers or reports that companies or researchers are publishing, industry research, you know, at Linear B, we do a lot of stuff of our own as well.

2:47And we just feel like we don't have enough time in the typical episode to give some of this stuff the space that it needs. So I think this will be a great opportunity. Starting out, it'll be probably just us two for a lot of these episodes. But over time, we want to bring in more expert voices to just bring greater variety and more fresh, up-to-date expertise that is breaking and cutting-edge news. So yeah, listen in if you want to get all sorts of new expertise from the people we bring on. Love it. And that's actually a perfect transition to our first story of the week because Anthropic reached out to us as part of their launch around CoWork.

3:26And I want to take a moment to talk about how impactful Claude CoWork is. This is a new feature launched from Claude that allows it to work in a way similar as Claude Code, but definitely more geared towards non-technical users. You have access to the folders to read, edit, create files, and give Claude a little more autonomy over your computer to do tasks related to knowledge work. This is a pretty exciting announcement. As of today, actually, it's available not just to Claude Max, but to Claude Pro subscribers as well. So be sure if you have a Claude subscription to go check out Claude Cowork.

4:05It's an ability for folks that maybe Claude Code isn't your best use case. you can still take advantage of the powerful dynamics of working with AI in this way. And a fun tidbit, too, I want to throw in here is that Cloud Code wrote Cloud Code work. So I think this is a sign of things to come. Ben, what do you think of this? Yeah, really cool to see AI building itself. Yeah, there was one line in the announcement that really caught my attention. And that was about how, you know, when Anthropic released Cloud Code, you know, they expected developers to use it for coding, obviously. And they did.

4:36but they quickly started using it for a whole lot of other stuff too. And that sort of opened their mind in terms of like how this could be applied. And that's exactly where my mind went as well. I was like, why does software engineers get all the cool new tools? Like, you know, us content producers like over here at Dev Interrupted deserve this stuff too. And, you know, we're, we're going to get into some really cool topics in this episode. And, you know, we'll, we'll talk a little bit about this Gastown article that came out. We are building our own Gastown, so to speak here at Dev Interrupted.

5:06That is to say, we're building these agents that help us develop and plan new content for this show. So, you know, naturally, as a content producer, we had tons of use cases that are outside of building software. And like the moment I heard about this, I was like, like, how do we start? How do we start applying this to content production? So yeah, this is a really cool thing. And I'm really thinking about like how amazing it would be to just have like a team of iterative agents, like constantly working to like improve old articles, for example, you know, or just like drawing from our rich history of expertise on our show to help us plan ahead.

5:40So, you know, I think this is really a great example of where this technology as a whole has the most potential to develop as a give it more capabilities rather than better frontier model performance, you know? Yeah, I think that's why it's so important for people to be paying attention to these new tools as they come available, because cloud stepping outside of coding and into more everyday tasks. And being able to work with AI in this way, applying it towards your knowledge work, allows you to capture the failure modes. Because failures aren't something that just happened and something deterministic like programming or writing code.

6:17They're also part of creating a really great copy or producing really compelling marketing campaigns or whatever your company or you might be setting out to do with knowledge outside of just coding. And this could be communicating with their coworkers on a technical team about the projects you're working on. There's so many ways that these things can go wrong. So by working with these tools, you can capture those failure modes and really find ways to improve them. So Claude Cowork is a way for people to kind of unlock those loops and those workflows for themselves. And I think it is like a step in the direction of Gastown for everyone else, right?

6:51Because I think you can't expect everyone to arrive at that eventual conclusion of being that level of orchestrator. But I think we can certainly ask a lot of folks who work with a lot of knowledge and documents to be able to work with them at scale and understand how things can go wrong and be literate about the tools they're using. And like I said, I'm excited to try this out. It was a research preview up until right before we came on to record this show, and now it's been released to all pro users. So I'll probably give it a chance to check it out. I'm a little slower than you might expect to adopt new tools like this, though, because if it's still a preview or an advanced version of it.

7:31Like, you know, I don't like to relearn software. So I'm hoping that, you know, I can maybe give a little more time to develop before I jump too heavy into it. But it's cool nonetheless. And yeah, so let's start talking about, so we've mentioned Gastown and we're going to get to that in a moment. But I think before we get to Gastown, we have to discuss this Ralph Wiggum phenomenon and the guests that we had on our show this week, Jeffrey Huntley. So Andrew, what do we have going on here with Ralph Wiggum? Yeah. So let's take a step back for a moment and talk about the Ralph loop and Jeffrey Huntley, its creator.

8:02So like Ben mentioned earlier this week, we had Jeff on the show to talk about the origin of Ralph and where it came from. and to you know ben and i we've been into you are many of our listeners and readers who have been part of the dev interrupted story this may have been somewhat familiar to you because we've been covering jeff's articles here on the show for a while for the past year ever since he started to stumble into how ralph could be created and then eventually used it to create his own programming language completely on its own just to prove a point right uh there was a lot that we all learned along the way covering those articles and so getting a chance to sit down with him and understand the RALF loop and how it's transforming engineering was a really great treat for us.

8:45And I want to recap some of the things that really stood out to me and that mattered. Things that really are important to keep in mind as this conversation develops, because we're talking not just about Jeffrey Huntley's RALF loop, we're also talking about Steven Yege's Gastown implementation, and then of course, evolutions of it. Like again, Jeffrey Huntley's new loom project, which is his ability to capture the RALF loop and create basically an environment where these loops can drive software development forward. It's a new dawn of software engineering, one where we create the factories that create the code rather than create the code itself.

9:24So Ralph became mainstream, I would say, probably in the last two weeks. It's making big splashes on the news. It became an official Claude Code plugin, allowing people to actually kind of utilize the very simple Ralph Bash loop to iterate over like a PRD or design document to get a desired output. And people have been proclaiming incredible gains from using this tool, being able to point it at things and wake up the next day with fully completed software. All of this, of course, has been open sourced and easy for everyone to understand and track. Like Jeffrey's whole point is to wake people up to the reality of how software engineering is changing.

10:06So if you're listening and you haven't checked out the story yet, this is your call to action to do so. I think it's really important that everyone pays attention to what's happening. And I think it's kind of like a nod of where things are going to go. There's a lot, again, to dive in here. I want to start there just by setting the scene for Jeffrey Huntley, Steve Vega's Gastown, and then, of course, Loom. but Ben you know anything you want to add at the top of this? I think at its essence the RALF technique is really just breaking tasks down into something that can have clearly defined success criteria and then having an LLM attempt to solve it over and over again until you're satisfied that it has been adequately completed.

10:46It's a very simple concept which is kind of one of the cool things about it but it's also you know it's kind of a double-edged sword it's a little it's a little inefficient too. But, you know, we've been following Jeffrey for most of this year, like you mentioned. I would even say that like a lot of our own practices within software development are defined by what we've learned from some of the stuff he did early in 2025. It's really great to see him taking that idea further. And, you know, it's almost like Steve Yege in Gastown and Jeffrey are sort of like lockstep and key right now, like working, And like they're in tandem in terms of like how they are maturing, like their own practices, but also like all of software engineering as a whole, it feels like.

11:29There's some great warnings that come with Ralph, of course. You know, you got to monitor your API costs because this thing will just run through them if you let it. But yeah, it's it's it's a Ralph is really just I think it's a new fundamental skill that we all need to become aware of and start to incorporate into what we do. yeah because like what you're what you're buying here when you get a loop when you're putting a loop like this in action is eventual adempotence like the loop is the hero now and it's your job to create a successful loop and the loop is still messing up in all kinds of ways but what we learn from jeffrey is that you know software development as an art as a as a career is dead it's dying this is this is what ralph is replacing but software engineering is more alive than ever because there's so many failure modes as part of building systems like this because now we're we're abstracting our roles as the builders one level further and like i said we're creating the factory that creates the code and so it's about understanding what's going wrong inside that factory um and it's about understanding that there's a new set of skills and we've been talking about oh there's a new set of skills i think on this show for like all the last year i think it comes out of my mouth all the time but now there's an even newer set of skills that are emerging around orchestrating this information.

12:49Instead of being a decision manager on one AI or a few, you're now a decision orchestrator across a fleet or a swarm of these types of tools. And that really changes how people are working with it. And you might be thinking like, oh, well, not everyone has to work with AI that way. You're absolutely right. But the engineers that do are unlocking incredible gains that it's going to be impossible to ignore. and that rift is only going to get bigger, I think. So those are some of the things I want to call out about the Ralph loop. You should definitely be paying attention, but I wanted at this point kind of move on and definitely kind of talk about Gastown because Ralph, if it's the inner loop, then we want to talk about the outer loop and that's what Gastown is.

13:33Yeah, exactly. Like Ralph is almost, you know, it's like the atomic unit of how this machinery now works. Gastown is what it looks like when you start building a machine around that atomic unit. So why don't you walk us through what this is? Because we've mentioned it a few times now. And I feel like to really fully understand the implications of Ralph, you also have to understand Gastown. And in my interpretation of Jeffrey Huntley's loom is it's almost like his own version of Gastown, but he's also solving different problems. But why don't you walk us through what Gastown is, Andrew? That's right.

14:06So what happens is when you work with something like a Ralph loop for a while, where you have a plan of what you need to build. You set your AI after it with a very clear and actionable set of tasks. And then it loops over and over again in a simple bash loop. Again, people, that's all Ralph is, just to demystify it for you, is a bash loop. And once it achieves its success goals, it exits. So imagine having a bunch of those running in parallel or running in sequence on all sorts of software engineering tasks, not just creating software, but reviewing it, merging it, creating tests, doing cleanup, or doing the task management.

14:44There's a lot of different roles and hats that are played in software engineering. So this is about going one step further and creating your software engineering organization. It's about orchestrating the inputs and outputs across your agents to have a code output of an entire team yourself. And that's ultimately what the Gastown experiment is about. Gastown itself is a bunch of Ralph loops with very specific kinds of roles that operate in specific ways. What CVA is doing here is he's exploring the complete unknowns. He's in the fog of war and reporting back on what it's like to combine these systems and give them these identities and have them combine and pipe this information around in order to build software.

15:30where he's discovering the failure moat and capturing them in Gastown. And I'll say at the top that if you read the Welcome to Gastown article, it's impossible to just not fall completely into this world that he's built. And it's kind of intoxicating to think about how many metaphors and names he gives to the things along the way, to where it almost borders on the preposterous. You have polecats and slinging threads and beads of work and meows, which are molecular units. There's a deacon and a mayor. And in his most recent, you know, operating manual, he talked about some of the failure modes where things go wrong.

16:07Like, I think it was what the deacon was murdering tasks. So there was like a murder mystery going on in Gastown. So like, by personifying it, at first I was like, you know, it's confusing for me to do that. But I really love that he personified it because it actually gives you something to latch on to and understand as it evolves. yeah and you know what role i think he he missed is the chronicler that like writes stories about all the things that happen in gastown i feel like that's what your role is as the operator of gastown because the the neat thing about gastown is that it operates as like your you know you're at the very top you operate with your mayor right and so you you basically communicate with a top-down orchestrator who handles all of the day-to-day and the other stuff underneath so uh that's the real key of a software engineering org think about the input of what goes into it like your c-suite needs something from the engineering team they don't go down all the way in the trenches and tell all the different people what they need to build and do in order to do it they tell the person at the top and they make a plan and they implement it that's what gas town is it's trying to capture that system in one place yeah and i think at the core of it i think the the point that yegi is is really making with with this is that you can think of clawed code as being like basically just a building block.

17:22And what we're really missing at this moment is like the orchestrator or the Kubernetes for agents, as he describes it. And yeah, there's lots of warnings on this thing. I think he says rule number one is don't use Gastown, or at least if you're not at a certain level of, he also introduces a really interesting model of agentic development skill level. You know, you can actually sort of like graze yourself on your ability to wrangle agentic development. And so I encourage everyone to definitely go check that out too and just figure out where you are actually relative to this. I was successfully warned off.

17:57I don't think you were, Andrew, because when he was describing the people that could potentially leverage something like this, I was like, oh, that's actually the level that Andrew is at right now. So I don't know, Andrew, have you built Gastown yet? I haven't built Gastown, but I've certainly experimented and tinkered with the ones that exist, including this Gastown and the one we're going to talk about in the moment just for a brief moment at the end of the story about Loom, just to round things out. I think what I'm learning from all of this is that, you know, they're all screaming the don't use this label.

18:29And I think that's kind of like a paradox of somewhat in open source to open source or technology, but then compel people not to use it. But obviously it's a viral project. and it's like kind of like a rite of passage i think to look at this and understand how it's working and why all those parts even exist um and you really should be using it getting inspired from it figuring out how it breaks and then ultimately making your own there's tons of forks of gastown and already of loom and i think it's about discovering what your own is going to look like there's a lot of simplicity of the boil down out of the metaphor so this is your chance to you know, get inspired and make your own.

19:07Because if you do that as an engineer, as an engineering leader, and you can operate on this type of level, you're going to set the playbook for everybody else. Yeah. And you don't, and I feel like it's really important to understand you don't have to build all of Gastown in one day. There's actually a learning process to this that we all have to experience. Like there was a concept that was proposed in the emergency manual of heresies, as he called them. These are effectively like bad practices that sneak into the code base, like as these agents are operating. And you need mechanisms to root out those problems and fix them.

19:46And this is where like that, you know, you mentioned the molecular unit of work that was, that Yegi proposed or a meow. And I think that's where this, where some like concepts like that become really important. And if you can start to solve for that concept itself, you can then start to construct Gastown. And even just solving that problem itself can really create a lot of benefits in the short term for you. You know, when I think about it, like, you know, LLMs have a higher probability of reaching success when the task that you give them is smaller. And, you know, if you give them bigger tasks, there's a risk that these so-called heresies get introduced.

20:25And, you know, there's just lots of practices like that, that you need to be aware of just fundamentally how LLMs work. But yeah, I will also mention, another thing that really impressed me about these articles is that he's clearly applying these techniques to image generation too, because the images he's doing are very well crafted. They're all AI images, but they're very well crafted. And I suspect he's got some sort of orchestrator for building prompts for those images. So I'm actually really impressed by that. It's almost like there's like a photographer agent in Gastown. who's walking around and taking pictures of what's going on and that's what we're getting.

21:02Yeah. But I think this is the year. This is the year that we all start implementing these self-reinforcing AI loops. Ralph and Gastown, they're just the first step. And honestly, whoever figures out how to productize what Yegi and Huntley both are describing or demonstrating, there's so much potential in that. It's going to be an incredible opportunity when we have productized versions of this hitting the market. And if you're listening to this and your head is spinning, don't worry. You're not alone. This is a brand new developing story. We're on the edge of it here covering it. And that's what was so great about having Jeffrey here earlier this week, because this news is just starting to break.

21:43The cat just got out of the bag, people. So don't feel like you're behind. Really, we're here to really Paul Revere running through here like, oh, it's coming. It's coming. Like we're just coming to get you ready as our listeners, because, you know, the engineering world is certainly changing. And speaking of how it's changing, I want to talk about how some of what we talked about is going to impact the role of the individual contributor. Because I think that's the big lingering question in the room, right? That's like the elephant in the room, really, is that if you can run a Ralph loop, you can have, you know, days or weeks of output of a pretty seasoned software engineer.

22:19year and you can do it for, you know, Ralph Loop running is about$11 an hour on API electricity consumption costs. That's, you know, less than minimum wage. The unit economics have changed. So this article that we read also about AI killing off the individual contributor, there's a lot of nuance to it. I think this is a take that has come up a lot in the last year, but I think that especially now, there are some more roles that have shifted. So Ben, And what were some of your initial thoughts about what all of what we've talked about today kind of impacts like the IC and their role in an organization?

22:56Yeah. So the core tenet of this article is that AI is killing the individual contributors. So it's essentially arguing that the advances in AI have shifted the role of software engineers from being less like an IC, more like a manager. So this requires them to oversee and coordinate multiple AI agents rather than performing direct coding themselves. And the author even describes how daily tasks for engineers are now more focused on setting priorities, choosing architectures, and then matching tasks to the strengths of various AI models. There was a really great quote in this article about how a great day of management isn't saying that I laid 350 bricks today.

23:39It's saying that I helped unlock a person who laid 350 bricks. And that's really what it feels like when, particularly when you get into this agentic style of development, it's about unlocking your agents to do work on your behalf. And it's honestly, this article is a really great summary of how I have felt as my LLM usage has surged. You know, I think really what is happening is hierarchies are flattening because, you know, when you're working with AI, it is kind of like working with somebody who reports to you, but the feedback cycles are immediate, which really changes the dynamic of it. So, yeah, this is definitely worth reading as a reflection.

24:23I don't think it's too surprising what is being said in this article, but it is definitely worth reflecting on to understand if the way you are engaging in work today matches with what appears to be around the corner. So, yeah, what do you think about this, Andrew? Yeah, like I said, I think this is a take that a lot of people have had in the last year, but it comes and it hits a little differently this time in light of everything we've been covering recently. And I think it really does call out the role of somebody evolving to being an orchestrator, somebody who's managing a lot of context, and being able to communicate it clearly both down to their agents, but off to other humans.

24:59there's a lot of different handoffs there of being able to work with tools in this way that we're still figuring out but ultimately i think that the value of an ic it drifts from pure output to taste and aligning decision making with impact for the organization and if you're already a developer or an engineer who thinks in that way you're already in a great position to utilize these tools to have more impact and what you do every day is just not finding the you know your synergy with the tools that are going to work for you. I think it's like, you know, you get people who used to be 10x engineers, like now I think you're going to get people who are like 10x orchestrators in terms of how they're able to parallelize, you know, parallelize like in sequence, like lots of work and be able to review it.

25:47The folks who develop systems for that are going to just be massively productive. All right, let's round it out with a fun one. We've been talking about AI this whole time. Let's talk about, I don't know, maybe it was built by AI, but it certainly is not an AI story. Creepy Link is this link that, this creepy link that our producer Adam sent over to me as he often does. So yeah, named Creepy Link, one word. It's a URL shortener designed to intentionally make links appear suspicious. And it's a joke service. You know, there's disclaimers on it. So don't feel like you're, you know, they want to be clear.

26:22This is not intended to be seriously used. It's not bit.ly. Yeah, exactly. And I did play around with it, and they are very creepy and suspicious links that they generate. So, yeah, what did you think about this, Andrew? Because I had a good laugh about it. I cracked up with this one, too. I think it's fun to play with that dynamic like this of, like, oh, you make the link look creepy or unsafe. I think it teaches us about our own expectations. Like, what even makes a link trustworthy in the first place? None of these so-called creepy links are actually dangerous. They look massively so. But there's lots of links that we click every single day that are dangerous, that don't look creepy, right?