# Agentic Collaboration and the End of Cheap Code

**Podcast:** Dev Interrupted
**Published:** 2026-03-06

## Transcript

AI agents everywhere.
I'm I'm curious, Andrew.
You got your hands on some of my agents in the last week.
How many hours do you think they saved you?
Oh well, your agent saved you actually a lot of hours this last week.
Uh, because they were super useful for the tasks that we had at hand.
But I I maybe also that same question back at you.
Uh, you know, how how how many hours of my agents saved you in the last week?
Uh well in the last week, nothing, because I've been using my agents this whole time.
It is really fun how we're just like ping-ponging information back and forth between our agents at this point.
Like they just grab whatever they can, go over to our little laboratory we have for them, and then like do what we want, and then eventually like the signals go back to the other agents, you know.
It's pretty fun.
It's pretty funny how quickly I was able to change anything that you handed off to me as well.
Like to the point where I had to like bring in a whole dedicated agent just to figure out how your project and my project go back together.
And I gotta say, this is still one of the biggest unsolved problems.
And I think collaborative agentic engineering.
I experienced this last year when I was on Code TV's episode for Goose, and we did uh it was a two-person challenge vibe coding with sub-agents.
So you have two people using teams of sub-agents, and then at the end I had to combine my front end with someone's back end, and that's you can watch you can you can watch the you can watch the video to see how that turned out.
It's uh a pretty hilarious saga.
Um yeah, we need to go from like one to two.
It's like one person with agents to two people with agents working together.
Like, how do we actually do that?
Which I think we're gonna cover today.
So let's just get into the Friday deploy.
Uh I'm your host, Ben Lloyd Pearson.
And I'm your host, Andrew Ziggler.
Yeah, and here's what we're covering today.
We have open claw rocketing into a risky orbit.
Steve Yeege back unveiling the agent wasteland, which is what I want to talk about, uh related to what we were just bantering over.
We have perplexity launching personal agents, and then dev work falling below minimum wage.
Andrew, let's just start right at the top and talk about open claw, because I feel like not a week goes by at this point where we we something doesn't hit the news about it.
So what do we have here?
Yeah, so open claw rocketed to the top of Mostard status um on you know star charts, which is the famous way of seeing how quickly and virally uh projects get adopted.
It rocketed up to 250,000 plus, you know, stars within really just a few months.
It's one of the fastest growing projects ever on the platform.
Uh I think it speaks to the the viral uh like takeoff of this project, the fact that it has definitely left the original audience that it was built for.
Uh, and it's kind of staggering to see that huge number.
But what do you think of this?
Yeah, so to be clear, this is GitHub stars that we're talking about, and open claw just this phenomenal growth to 250,000 stars in like four months.
Uh, which I I looked at the leaderboard this morning, and I guess that puts it just outside of the top 10, which is pretty pretty crazy considering that how long many of those top 10 repos have been around.
And a lot of people are like comparison comparing it to Linux because like it just surpassed Linux, uh, like the number of stars that Linux has.
But I I don't honestly don't think that comparison is really fair because like if you know anything about Linux, like you know that Linux development still is happening over on git.kernel.org.
And the GitHub repo is just like the it's a simple mirror.
Like there's no community activity that happens there.
So, you know, I I've I've wondered just with the with this news, um, if Linux was birthed in the GitHub era, like would it rival this the current success of open claw?
So uh, but you know, I think really what I take away from this is that it's just ongoing proof of this prediction that I've I've had that the biggest AI benefits that we're going to achieve in the near-term future uh will come from improving how LLMs are applied to practical everyday situations rather than better and better foundation models.
Like I just like foundation models are great, but like, you know, people would just need agents that work with the tools that we work with every day.
It's funny, Ben, to you know, hear you talk about all the caveats related to the repo, because I've also seen a lot of posts about you know delineating between this and like oh, is it compete compete with like the the archival repos, which have like more stars than these.
It reminds me of a character on Party Down that really wants to open a salad chain restaurant.
He is really excitedly telling everybody that it's the fastest growing, non-poultry, non-coffee franchise in all of Southern California, and it's just putting all of those caveats on it, just makes it extra fun.
But this is just also a reminder that GitHub has always been a social network, it really speaks to the viral nature of the platform itself and the ability for it to bridge from engineering into the rest of the world.
And I think this is the clearest sign yet, like I said, that open claw has left the building.
And what I mean is it's not just for engineers and techies anymore, like your bus driver and your aunt are starring this thing.
And I think this speaks to a general hunger for multi-purpose agentic assistance.
Uh, but they're with any kind of project that has massive growth like this, there's so much potential for security and abuse.
And so uh unfortunately, a lot of this is gonna be discovered in fixed and prod because it's been so widely picked up and adopted.
So definitely be reminding your aunt to upgrade her open claw as those things get patched, because I think that's gonna be the biggest danger ahead.
Yeah.
Well, I hope my bus driver's not operating open claw while driving the bus.
But you know, if you yeah, I mean, well, they should be giving the orders before they start driving the bus and then have it working while they're driving the bus and not having to look at it, of course.
Fair.
Um, but you know, this really kind of feels like you know, the article we're gonna share in the show notes, like it it highlights how it achieved this phenomenal growth despite having these like significant security risks and other problems that it creates.
And it feels like a very like common story in tech of like three steps forward and two steps back.
Like technology advances at like some pretty substantial margins, but then it breaks a lot of the fundament fundamental practices that we've built up over time.
So we have to like spend time like getting back to where we were in terms of like just raw capabilities.
So, you know, one example of this is like when smartphones started to take off, like even before touch screens were a big thing.
You know, everyone was building web apps and or they're building apps for phones and building websites to work with smartphones.
Uh and there was this period where I felt like the entire like web and app experience just got like substantially worse than it had been for many years because we were having to reinvent a lot of things for this new era of technology.
And of course, I mean I the parallel between mobile and uh AI, I've heard a lot, you know.
I think generally speaking, people who have been through both are they feel very similar, even though AI does feel like it's a it's a bigger scale than mobile ever was.
But you know, it's it's very clear that these AI agents can break all of our security conventions in very short order, they can do it at scale, they can put you at risk really quickly.
So I'm still in the camp of like open claw being this like incredibly cool but also equally terrifying experiment.
Like it's it's just it's both.
I agree.
All right, well, let's move on to the wasteland or a thousand gas towns.
So we're talking about the newest article from Steve Yeege, who we continue to cover over here at Dev Interrupted because we're learning so much ourselves from him.
And Yege announced the wasteland, which is a federated workflow that links thousands of gas towns via shared wanted boards, and it has a system for moving work through states with reputation and multidimensional stamps and data that's stored in Dolts and its versioned and auditable.
Like it's just taking sort of the the brilliant chaos of Gas Town to the next level, you know.
And Andrew, you and I have been talking about this concept for weeks now, really ever since we learned about Gastown and started applying agenc work into our own uh days.
Uh and we both like I I think clearly understood that like it doesn't stop at level eight of Yeege's model of agentic development.
Um so we've been like wondering like what's next, like what's the level nine, the level 10?
And you and I have been having a lot of one-on-one conversations lately about like what it means for multiple people to have orchestrators.
I mean, that's how we opened this conversation today.
Uh and what does it mean for multiple for multiple people to have orchestrators that work together to solve problems and coordinate issues?
And you know, and Yege's back with this this wild and interesting metaphor that fits like his typical flavor of storytelling.
Uh, but it really does have like my brain churning because it like the wasteland metaphor itself is like both kind of hilarious and brilliant to me.
Like, I I like to imagine that the orchestrators that we build, you know, these agent orchestrators that you and I are building, uh, they're they're like basically creating this wasteland of finished work, like all around us, right?
Like any task that we build an orchestrator for, that task is suddenly solved whenever we need it to be solved.
Like there's no, there's no wait time.
We're just waiting for the tokens to process effectively.
Uh so you know, if another orchestrator wants to come and work with me, it makes no sense for it to try to to work in that wasteland.
Like it has to go across the wasteland, like bridge the the chasm, so to speak, and learn how to navigate those vast expanses of these like complex finished work environments, and then dock with my orchestrator in a way that they can work together.
You know, that's kind of how I'm like visioning it myself or what we're doing.
So I wonder if that lines up with how you're feeling about this, Andrew.
Yeah, I think that gas town evolving into the idea of the wasteland, that concept makes a lot of sense to me because what gas town solves for the individual is it lets you split up work into atomic units and delegate it into specialized agents or workflows to get done on loops.
And so the natural like aggregate of that, if you were to go one level higher, is to have them working together almost like a symphony, right?
Having different types of gas towns that do different types of work that can coordinate together, almost like how within those sessions they would be coordinating specialized sub agents.
So that's the really the brilliance, I think, of what uh stands out here.
Yeah, guys, obviously have been many steps ahead of all of us this whole time.
So whenever a blog drops from him, I scour for clues because it shows where we're gonna put our feet, where we're gonna climb up that mountain.
It really leaves the footprints that we can follow.
So some things that stood out in here that I'm definitely going to be trying out in my own orchestrator systems and flows, uh, one of them being reputation.
I thought this was a really smart takeaway.
Um, the idea of having a record, a a core system of uh that's an accurate record of the work that needs to be done, who did it, and then the level of accuracy and completion at which they did do it.
All of that exists within a ledger.
This can be something that's almost like immutable.
You start thinking about something that is storing the work that all of those agents are doing, improving it in a shared space.
And the value you get of this is coordination, you know, uh it other types of technology, especially things like and and like blockchain technology as well.
That's how they operate together.
And so really he's borrowing a lot of concepts there, reputation being one of them.
But what stood out to me as well is the idea that you know, agents are not the blockchain.
They are accruing real specializations and domain expertise around all sorts of different nuanced work.
And so being able to accurately and reliably judge that work and if you can trust it is how I would be able to let my gas town or my specialized agents in my gas town work with yours, and you'd be able to trust that.
Another part of this too that really was interesting to read about is the federation itself.
Because like two months ago, I made a joke on here in one of our news segments, Ben about like I was like, oh, what's gonna happen is like my gas town gonna call your gas town.
And now, like egg on my face, oopsie, I was definitely wrong because it's a wanted board.
That is where this is going.
Wanted board, you know, telephone, however, you want to dress it up.
It's it it answers to a almost like uh one level higher than like the A to A protocol of like how to like the how does this mass of agents and this mass of agents work together?
Yeah, um, and both of these ideas together, the reputation and the federation, they become a natural framework that enable people to systematize and combine this work together in systems that obviously you know don't have uh like a mutual understanding of of trust.
Obviously, within like your own company, you probably don't need this level of granularity in order to interop with your co-workers.
But when we're talking about you know the wild west or out in the opens, or in this case, wasteland, this is the kind of system you need to protect yourself.
Yeah, and I mean Yage, you of course you gotta always remember he's working with a with a large open source community at this point.
So that's a very different, you know, you know, it's it's hard to build trust in that environment.
You have lots of untrusted actors in that environment as well, you know.
But there was there was one element that really stuck out to me that I absolutely love, and that's the RPG metaphor that he has towards the end of this article, where he had like a little like character sheet that he had produced that sort of sort of shows like the skills and the the reputation, as you mentioned, of the the orchestrator.
And I I think we need to steal that, Andrew.
Like, we're gonna steal that idea.
All of our all of our agent orchestrators are gonna get their own character sheet.
Uh, and yeah, they're gonna have bounty systems, wanted boards are gonna have all that stuff, I think.
Amazing.
All right, let's talk about this perplexity computer.
What what do we have here, Andrew?
Yeah, so this is uh another side of a similar coin coin: the idea of how do you, you know, combine agentic systems together to work reliably and at scale.
This is perplexity announcing the perplexity computer.
And it's a general purpose digital worker that can create and execute whole workflows and do long-running jobs that can run for hours or days or weeks or months using a very intelligent multimodal orchestration system.
And what makes this different from, you know, other harnesses or operators is that the perplexity computer is operating more on a granular level, uh almost operating on the operating system level itself.
It speaks to some of the natural evolutions of what technology has to do to meet agents where they are, because I've said on the show before that, you know, agents are arrived in internet that's not yet built for them.
And I I love projects like this that reimagine what that compute could look like for them.
And it's not a foreign idea to the show either, because on our newsletter, Lenny Press last year of Amplify Partners, he rose the wrote a really great guest article about programming languages and how they should be designed for agents.
Uh just recently I sat down with Matt Boyle at ONA about how they're designing cloud environments for long-running agents in a similar way and protecting uh the rest of the technology stack on a kernel level in terms of what the agent can do.
Um, there's a lot of benefits in this and complications, but one thing I think is for sure, it's um operating and restricting and redesigning at almost the kernel and operating system level is something worth exploring.
And I think this concept is only going to get more uh proven as time goes on.
So when when I was reading this, it felt like this is almost like perplexity's answer to like clawed cowork or open claw.
Is that is that your interpretation of this, Andrew, or is it something else?
It seems to be.
It's the idea of the answer engine.
How does that evolve into its natural next state?
It can take those answers and then it can turn that information that you're scouring long term into systematic reports.
You know, this there's also a technology like this called Scout that I've seen as well.
And these are long-running like um a search agents.
You basically set them on some search queries or some like certain kinds of like uh things that you want them to constantly stay on top of.
They search on a regular basis, go deep on articles and prepare reports for you in the background.
And the ideas they run for weeks or months, just like this, you're designing and executing the workflows with like an answer, like a long-term answer that you can't get from one query.
Yeah.
I feel like this space is like getting competitive about as quickly as like the foundation model space has gotten competitive, you know.
And uh, you know, I'm I'm not personally a perplexity user.
Um, I have used it in the past, but I have always been fascinated by what they do because I think they they do often take very creative approaches in this space.
You know, and I'm just thinking back to like when they first hit the scene, you know, everyone back then was basically a chat bot.
Like that was all of the GPT experiences that we had out there.
Uh, and perplexity was really the first one to take this approach of like applying that to like create a new search engine for the web, which is a really interesting application that's now starting, like everyone else has sort of caught up to that and done their own version of it.
And they've also done a lot of interesting innovations around like content generation, collecting feeds of information.
So yeah, yeah, it it totally makes sense that they continue to innovate in this space.
And you know, for now, Lo, I feel like you and I, our team, we've been taking really the approach of like building our own agents mostly rather than trying to get something out of the box, which is kind of what my impression of this is a little more, uh, which is why we've stuck with like clawed code, for example.
Um, but I think if you know, correct me if I'm wrong, but it seems like if you were if you're someone who wants more out-of-the-box experience, like this may be something that that would benefit you.
Does that seem right?
Yeah, if you if if you if you want to operate with long-running answer engine agents as a service, then perplexity computer is probably the something to look for.
Of course, if you're orchestrating and collecting that knowledge on your own with workflows in the background, maybe you're this is already covered.
Uh but this is definitely gonna this is definitely a sign of like more types of services like this to come.
Yeah, I'm I'll definitely give them credit because I think they are on the right thread in terms of like, you know, again, the types of the types of things that we should be building to to get better more out of AI, you know, so it's a pretty cool project.
All right.
Well, let's close out our lineup with uh this story from friend of show, Jeffrey Huntley, about how software development is now cheaper than the cost of hiring a minimum wage worker.
What do we have here, Andrew?
Yes, this is um the scoop on software development and how that practice is dead compared to software engineering.
This is Jeffrey Huntley's core operating thesis.
Uh and he's famously written that software development now costs 1042 an hour because that's how much it costs the run of claude code session on a loop on a virtual machine.
Uh and this is coming at a time where when you go to any kind of tech meetup or even AI meetups now, it's oftentimes full of full of people in the room that aren't traditional engineers coming from all sorts of different backgrounds, embracing the tools.
I think this is like uh really speaks to the widespread availability of it, but it also speaks to the almost like uh immediate devaluation of a specialization, where if your output was just writing code, then you know that that time has changed.
And that's really what Jeffrey gets at in this article.
Um this was definitely making the rounds on LinkedIn as well.
Yeah, and this was something we taught we spoke to Jeffrey about when he came on the show uh a few weeks back.
And you know, I first of all I want to say like, you know, we don't mean to degrade anyone's work by talking about this subject.
Like, um, if you're a software developer or if you're a minimum wage worker at like a fast food place or some sort of service sector job, like I don't want to minimize the value of those people or the roles that they bring, but I think it is really important to understand this or to point out how software development is something that like used to be viewed as a way to achieve like relatively good personal wealth.
Like you could learn to write code and make good money off of that.
And that simply isn't the opportunity that's available anymore.
And I think it's really important because there's a lot of nuance here that that I think often gets lost.
And I think definitions are incredibly important for this.
So software development is just the act of turning requirements into code.
Like you're basically paid for understanding how to write code, as I said, versus software engineering, where you build the requirements and understand the users, the architecture, the technical decisions.
All of that requires higher order capabilities and skills.
So it's the former that's being replaced.
So software development.
And that's being replaced wholesale by AI right now.
Uh, you know, and frankly, we're seeing the same thing happen on the content production side, what we're doing here at Dev Interrupted.
You know, you and I have been compressing workflows that used to take us days down to hours or sometimes even minutes or seconds.
You know, like just last week, we were talking about our how the the all the time we save with the agents that we built.
Uh, you know, we built this agentic system.
We probably saved 40 plus hours of work in the last week.
That's like an an extra week's worth of work that we compressed down to hours.
Uh and this was a project with Tide DevLite.
And so it had, it made, it made a massive impact on us.
Yeah.
And I appreciate the distinction that you make between software development, software engineering.
I think it's crucial to make, but I also think it's a reminder to folks who are tr software developers that you are also a software engineer.
You know, you have the skills behind scoping retirement or making requirements and understanding the technical decisions and making the architecture, those are all latent if they're not expressed explicitly.
It's just a matter of finding them and being able to utilize those skills with things like agents to do maybe more of the work that used to occupy your time before.
If you can make that jump, which I know and I'm fully confident any software developer listening to this or otherwise is capable of doing, then you know, you're gonna be well positioned to ride the wave.
I think the the important thing here is to you have to take action because the gap will only widen.
And the longer that you wait, the harder it will be to cross the chasm.
But I like I said, there's gonna be a lot of new works, uh new work for engineers and new systems to be built and combined.
It's an opportunity to build and define skills before anyone else.
Going back to the story of Gas Town and being a highly specialized, highly reputable person capable of outputting tons of code.
Imagine that is your next thing to optimize for.
How can I create an agentic system that's hyper specialized and something that I have a lot of domain expertise in and then own it and its outputs?
And, you know, if if you're not hungry enough for that kind of chase, then I think there's somewhere in the middle that you can fall, but you do have to find that place for yourself.
But just keep in mind that like AI is moving so fast.
So it's important to stay on top of each of each and every week.
And it all brings different things.
That's why we're here every week talking about it.
But on the upside too, I think we're gonna be in a world where engineering proliferates everywhere.
All companies become technology companies in a way that they weren't before.
And you're going to see a lot of really cool new engineering-related roles that simply weren't possible until now.
And you could find something that could end up being your life's passion as a role that simply doesn't exist yet because the world is evolving into that.
And if you want to like brainstorm or explore some of those, I really recommend uh Scott Werner of Works on My Machine.
He's we've covered his articles here before.
He has a really cool uh system called the Traffic Jam Explorer.
It's a clawed artifact that naturally explores the evolution of ideas of like you introduce an agent to an industry or a problem.
How does that beget this?
And then if that begets that, then what becomes of that?
And it's basically a philosophical tree of potential futures.
So it's really cool to dig through there and find maybe unique opportunities for yourself that could become the reality tomorrow.
Yeah, I actually have not read that article yet, so I'll have to go check it out.
But I love how this episode really is coming full circle today.
Like it feels like everything we're covering is like very interconnected, you know, because I think this is where the wasteland metaphor really starts to come back.
Like our agentic systems, they effectively allow us to expand the breadth of our scope within our roles.
So we can do new things, we can take on new challenges, we can solve bigger problems.
And we're creating this sort of like bubble of problem types that have been solved around us, you know.
You know, again, getting back to this wasteland, like all of these tasks that are around us that just are solved by default now because we have agents that do it.
And, you know, I think the like Jeffrey brings up a lot of really great opinions or in this article, particularly around like seat-based pricing.
Like that is a major risk for a wide range of reasons.
If that's if your company still depends heavily on it, um, the best pricing models right now seem to be things that are usage-based because that reflects the utility and the value that the product brings to someone or an agent for that matter.
And there's a lot of tough challenges to grapple with right now.
But like you said, there's an extraordinary amount of opportunity as well.
Like if you're someone who's out there who's really starting to take advantage of agentic development or agenc work otherwise, and you feel like you're thriving up because of it, like you should embrace that and like really push into it because I think it's an incredibly valuable skill to have in this moment.
And, you know, just from personal experience, I know you and Andrew, I feel like we're having this almost like leapfrogging effect where where I'm able to work with my agents to get a breakthrough, and then your agents come in and they get their own breakthrough, and we sort of like uh uh we're accumulating all these benefits at a rate that we've never been able to do it before.
And because we're we're both operating uh in this way, like the the sum of our components is is or the sum of our whole is greater than the component parts.
I think that's how it goes.
Something like that.
Yeah.
But I mean, you know, long story short, collaborating in this agentic way, it's a lot of fun.
You can do a lot of really interesting and amazing things that you didn't really expect you could before.
It's just a matter of learning the skills, and and now's the time.
Now's the time to do it.
Yes, collaborating is very fun, also be very chaotic.
Um there's lots of strategies involved.
So, you know, the wasteland's only the beginning of that saga.
I'm sure we're gonna see lots of really interesting scenarios that come out of that reality, and we're gonna cover them here on Dev Interrupted because this this world's moving pretty fast, but uh there's a lot of stuff to cover week after week.
Cool.
What are your agents working on this weekend, Andrew?
This weekend?
Oh, let's see.
Well, going back to the perplexity computer idea, I've been actually experimenting with the idea of creating almost like a kernel for an LLM to use.
I've had some prototypes I've gone back and forth with, and maybe this weekend will be uh the weekend I bite the bullet and try to build it out.
What about you?
Yeah, it's amazing how we think about these challenges at like totally different levels because you're like, I just need a new kernel, obviously.
And I'm over here, like, how do I get my agent to help me make sense of my life?
Well, I created this like scripting language that then I was trying to, I was thinking maybe I would have a fun experiment of maybe I could fine-tune an agent that was really good at using this custom scripting language that I had made because that's one of the fun parts too is that you know it's so easy to make new languages now.
I've already made one.
Our friend Jeffrey on the show, he's made one as well.
We covered Cursed back here on on the show back like last uh October.
Yeah.
Uh this is a C compiler.
And it's like uh I love having these kinds of utilities a C compilable uh and making them.
I I just wish that more people would uh adopt and use things like cursed.
So if you're a if you're a listener and you're building stuff this year, maybe maybe cursed is the project to pick up.
Yeah, yeah.
And I know Jeffrey mentioned something about wanting to get it in the Stack Overflow survey this year.
So, you know, Aaron, I know you listen out there, friend of show.
If if you can make that happen, hook us up.
We'd love to see it.
Everyone else might have to.
Yeah, otherwise we might have to kick off a writing campaign.
Who knows?
We'll see.
Well, they're definitely gonna be a writing campaign.
I think everyone else, this is your call to action.
If you're using agent agents this year to do some coding, maybe try to have them write it in cursed and fill that in on the developer survey this year so we can make sure that our friends have Stack Overflow know.
Yeah, absolutely.
All right, well, thanks for joining us, and we'll see you next time.
AI helps your developers write more code faster.
Here's the problem: your review process hasn't sped up.
The queue grows, reviewers get burnt out, cycle time stalls.
Linear B changes that.
Our AI reviews every PR the moment it's created, catching bugs, security gaps, and performance issues before humans get involved.
It even writes the PR description automatically.
Your reviewers spend less time on first-pass problems and more time on architecture and business logic.
Break the bottleneck.
See how Linear B accelerates your workflow.