LinearB's software factory spends more time on the spec than on the code. VP of Engineering Ariel Illouz spent the last year moving his team of about 30 engineers from vibe coding to spec-driven development, starting with a single spec-writing skill and a rule that every story above three story points gets a spec. This week on Dev Interrupted, he and returning guest Ofer Affias, LinearB's Senior Director of Product, join Andrew to walk through a project Ariel puts at six months without AI that shipped in roughly four weeks, at least two of them spent on specs. Ofer explains why proving the impact meant measuring AI adopters as separate cohorts. They close the conversation on where the software factory goes next, as agents take over more of the loops humans run today.
Transcript
(Disclaimer: may contain unintentionally confusing, inaccurate and/or amusing transcription errors)
[00:00:00] Andrew Zigler: So it must make it really hard to, like, do meetings in a sane way because your meetings get, like, split on two different ends of the day probably, right?
[00:00:10] Vincent Koc: Part of the
[00:00:10] Andrew Zigler: And so you get the... Yeah,
[00:00:12] Andrew Zigler: Part part, part of the fun
[00:00:13] Andrew Zigler: And, and, and, and there's probably different themes that are like, "Oh, this part of the day everyone's worried about this kind of thing.
[00:00:18] Andrew Zigler: This time of the day everyone's worried about this kind of thing." And you're like, "Guys, this is the same day for me, by the way." It's kind of a, a weird, a weird kind of world to live in. But it, it's also, like, the fun of just, like, I think working on a project that has global implications, right? That, that is so interconnected across the world and has a lot of voices contributing to it.
[00:00:40] Andrew Zigler: And, like, that's why I'm really excited to dive into some of the stuff that, that you're working on. Um, but, but really, like, uh, working on, like, a globally distributed project is quite a challenge I think in our world.
[00:00:54] Vincent Koc: It has its own plus
[00:00:56] Vincent Koc: and minuses, I think. talk
[00:00:58] Andrew Zigler: Yeah, for sure
[00:00:59] Vincent Koc: [00:01:00] experience in large scale open source work and, uh, large scale sort of like engineering work as well. Like being like a AI/engineering person for like an airline before and things like that. So like the scale and reach has been pretty,
[00:01:15] Andrew Zigler: Mm-hmm.
[00:01:15] Vincent Koc: pretty comfortable for me, but it's not something a
[00:01:17] Vincent Koc: lot of people are used to, so it's--
[00:01:19] Andrew Zigler: No, this is definitely when you have that kind of like high level
[00:01:24] Vincent Koc: stuff and learn the hard
[00:01:25] Vincent Koc: way, in my opinion, but it's
[00:01:26] Andrew Zigler: I think that's what makes it scary.
[00:01:29] Vincent Koc: Mm.
[00:01:29] Andrew Zigler: because when you're in a global world, it's never off. You're always breaking something. Even when you're trying to tiptoe as carefully as possible, you're still taking a sledgehammer to someone's reality.
[00:01:39] Andrew Zigler: Wow. Okay, well, you know, just diving in here, welcome back to our listeners to Dev Interrupted, brought to you by LinearB. And today we're diving into the story behind OpenClaw. We have a really exciting conversation in store. We have the chief architect of the OpenClaw Foundation, Vincent Koc and one of the major contributors to the [00:02:00] entire project.
[00:02:00] Andrew Zigler: And today we're going to really, uh, open the hood on the capabilities of OpenClaw, but also zoom in on the implications this has for, have for leaders within preexisting, pretty mature, large engineering teams, engineering teams that have engineers, which by the way, are heav- heavily leveraging and utilizing OpenClaw and OpenClaw-like technology.
[00:02:24] Andrew Zigler: What are the kinds of things that they need to be understanding to stay on top of the way that their engineering org is evolving? And I can't think of anyone better to dive into this with because you have so much, uh, insight into how folks are using this tool as such a major driver of it yourself. So, um, you know, Vincent, welcome to Dev Interrupted.
[00:02:44] Andrew Zigler: We're really excited to have you here.
[00:02:46] Vincent Koc: Likewise. Thanks for having me on the show
[00:02:49] Andrew Zigler: Amazing. Well, I, I wanna start off by first off congratulating you on OpenClaw 2.0, a major i- improvement and achievement as well to the entire code base in terms of the [00:03:00] features that have been available. We've talked about it a bit, a bit on the show because a really exciting part about OpenClaw, at least for us here at Dev Interrupted and, and some of the folks that have chimed in, is really about, like, the shared collaboration-driven experience.
[00:03:14] Andrew Zigler: This has been, like, the real hunger, I think, of engineers that have been experimenting with AI since the beginning of the year, since Gastown, you know, lit the world on fire and people realized that...
[00:03:26] Vincent Koc: The
[00:03:26] Andrew Zigler: Yeah,
[00:03:27] Andrew Zigler: B- okay, great. Well, you're saying beads. My producer might as well come off mute and just be like, "This conversation's over," because I say beads in every conversation, and everyone thinks I'm a crazy person.
[00:03:37] Andrew Zigler: I actually was just recently at LeadDev. I was giving the, the, uh, I was giving a talk there, and the talk was about why Miss Frizzle would be a great SRE. And really, it was actually just, like, a, a hidden lesson in how to use beads to just make everything possible
[00:03:53] Vincent Koc: absolutely love
[00:03:54] Andrew Zigler: time.
[00:03:54] Andrew Zigler: So beads are a fundamental part of what make this possible.
[00:03:58] Andrew Zigler: They're primitive, right? Uh, [00:04:00] s- Steve Yegge in his articles has talked about this extensively. Chris Sells is out there building with this as well, and... But OpenClaw, we've argued on the show, many have talked about, is also its own kind of primitive. That's what's so exciting about the OpenClaw experience and the way that it grows and it conforms to your domain expertise.
[00:04:17] Andrew Zigler: So, you know, what has, what has that been like for you to, like, develop and build something that, uh, that, that honestly has such wide use- usage? Um, like, obviously it has such a, a, a big perspective. So I'm just kinda curious to get in your head about how you think about these things.
[00:04:32] Vincent Koc: I think like for my personal expe- um, experience, it's been kind of interesting 'cause I've worked on large, large scale enterprise work from like, uh, airlines, banking, like I'm talking huge reach, but this just felt different. You know, I've worked on open source as well, but it's just the velocity of everything.
[00:04:51] Vincent Koc: I think I remember one stage, uh, we're at GTC NVIDIA, uh, last year or this year. This year, it feels like last year. [00:05:00] That's... You see what I'm saying? Like, even I'm getting it twisted. It's like just a few months ago, and Jensen is saying, "Oh, this is like the biggest thing since, you know, the internet or computing."
[00:05:09] Vincent Koc: I can't remember the exact words, don't quote me on it. But just the, the levels of insanity just kept increasing and increasing. I had like rap artists talking about this stuff on X. I had, you know, uh, C-suites of like major Fortune 100 companies like sending me messages of like, you know, issues they're having.
[00:05:28] Vincent Koc: It was just like the... It just kinda exploded like, it was definitely a hype moment there. I think that was def- definitely a hype cycle. But for the first time, it felt like agents and AI had like
[00:05:40] Vincent Koc: escaped the containment of like the tech ecosphere and the, the tech bro culture and all the rest of that that comes with it.
[00:05:48] Vincent Koc: And for the first time, it was part of the zeitgeist, you had like your Uber driver talking about this Claw AI thing. I remember I was like in a bar in, in New York in my hotel just like trying to take a breather [00:06:00] before this ClawCon event that we just had, and it was like people had like walked for...
[00:06:04] Vincent Koc: Peop- The, the queue was like three blocks long. um, the guy's telling me, "Oh, have you heard about this like AI thing? It's like this lobster." And I'm just like,
[00:06:12] Vincent Koc: "Oh my God, what is hap- what is, What is happening? What is
[00:06:13] Andrew Zigler: what is happening?
[00:06:14] Andrew Zigler: This, like, echo chamber of a world. What have I started? Yes, exactly.
[00:06:18] Vincent Koc: escaped, right? And I think
[00:06:20] Andrew Zigler: Yeah
[00:06:20] Vincent Koc: post that or like what I like to call the post-Claw era is like this like explosion of agentic AI that's one sort of like token usage has gone through the roof everywhere. But also like individual AI agents and things that people are building that are like somewhat like actually useful and like built around taking action and making change and doing something that's for you or for your company or for your team. And just seeing that explosion in the industry and seeing that all come alive has been amazing as well.
[00:06:50] Vincent Koc: I think it's great that we were sort of part of that, that story or that next chapter of what AI
[00:06:55] Vincent Koc: might look like for everyone.
[00:06:57] Andrew Zigler: Yeah. You're so right about it escaping [00:07:00] containment. It was out in the wider world. Uh, I... On this show, when we talked about OpenClaw and its meteoric, you know, uh, GitHub stars back when we could see those and, and just how many people were obsessed with the project, and also just how many people had downloaded it.
[00:07:17] Andrew Zigler: Early OpenClaw was just so massively downloaded, and people were experimenting with how to run it themselves. And for the first time, it felt like folks could have just a little slice of the AI fun for themselves because that's part of what makes OpenClaw really appealing is, you know, conforming it to what, uh, you're looking for.
[00:07:37] Andrew Zigler: Excuse me.
[00:07:41] Andrew Zigler: Excuse me.
[00:07:43] Vincent Koc: I think it's on.
[00:07:44] Andrew Zigler: um, I... From, uh, back- taking a step back for a second, like,
[00:07:49] Vincent Koc: Mm-hmm.
[00:07:50] Andrew Zigler: every- everyone is interacting with OpenClaw, downloading it, experimenting with it. You get a surge of interest, and this is something I actually really want to zoom in on because we've been, [00:08:00] here on Dev Interrupted, trying to crack the puzzle around what open source contributions and open source world looks like in an agentically enabled developer world.
[00:08:09] Andrew Zigler: Because in an agentically enabled developer world, the interactions between maintainers and contributors with ideas is totally different than it was when, you know, everyone was submitting code, and that was the main contribution at play. So as, like, an, uh, in, in this case, you know, the chief architect of the project, how do you think about organizing the way that people can come in and engage with the project and, uh, without, uh...
[00:08:35] Andrew Zigler: You know, that way you can kind of keep it on rails as a builder?
[00:08:39] Vincent Koc: I don't think there's any, like, perfect source to this. It's like a ongoing thing when it comes to open source projects and work, and I don't say... I don't think it's like a fundamentally, like, something I am personally, like, driving. I think it's a
[00:08:50] Vincent Koc: concerted team effort. But
[00:08:52] Andrew Zigler: Yeah
[00:08:53] Vincent Koc: when you think about a project like this and you think about open source, everyone wants to help, so you do end up with this, like, [00:09:00] this team of individuals that would, you know, essentially maintain the project. But then on top of that, you have that group of people working with the community and the individuals to figure out what's right, and I think that's where we've been always listening to the community. We have
[00:09:14] Vincent Koc: like a Discord of like, I think 170,000, almost 200,000 people, and I think we, we spend a lot of time on there.
[00:09:20] Vincent Koc: You know, Peter himself is con- continually on X messaging and responding to everyone all, all hours of the day, and I think we just engage, and I think the figuring out what people want, figuring out how people wanna help, and how we can slot people into that mix is something we kind of review on a continual basis, and it's like a concerted effort, and in some ways the community helps us
[00:09:40] Vincent Koc: decide what that, what that looks like.
[00:09:42] Andrew Zigler: Yeah, it's about sourcing the signals. Like, what you're calling out is really what I think stands out from a lot of, like, the heavy hitter, the large, uh, open source successful projects right now, the ones that are drawing attention to them
[00:09:53] Vincent Koc: Hmm.
[00:09:54] Andrew Zigler: folks, is really about c-
[00:09:56] Andrew Zigler: encouraging people to create the space for ideas, to come together and [00:10:00] share what works and create a space to collaborate.
[00:10:02] Andrew Zigler: Because when I think of just the, the widespread success and adoption and just mental share of OpenClaw within the world, so much of that, I think, has to do with the fact that there was a while where you could pivot left or right and find somewhere with somebody that you could pair up or duo up and figure out and learn OpenClaw together, and that's where I think mag- that's, like, the magic of open source, I think, is when people just spontaneously create this learning environment.
[00:10:32] Andrew Zigler: And I think that happened, um, around Op- OpenClaw especially, because it, it represented such an opportunity to experiment with what building and owning these things will, would look like. Which also brings me to, like, another interesting reality, is that, you know, for our ... So our, our listeners definitely are, are within engineering teams where, you know, having something like OpenClaw maybe introduces a lot of risk and questions, [00:11:00] and it's not a tactic that they could feel really like they can pull out of their tool belt.
[00:11:05] Andrew Zigler: Instead, there's learnings from OpenClaw, they get applied into how workflows are structured, into how we think about sharing context, and specifically with OpenClaw 2.0, the multiplayer aspect, because that becomes the most important thing that we're all seeking, is I'm agentically enabled, you're agentically enabled.
[00:11:25] Andrew Zigler: How does my Gastown call your Gastown? We have all sorts of fun versions of this right now, like Claw Tag and things where people are collaborating in Slack. And, and I really think that OpenClaw also becomes a place like this. So I'm curious, like, what are, what have, has it been like, uh, for you to interact with larger teams or, or teams l- of people that are using OpenClaw together?
[00:11:49] Vincent Koc: Yeah, I might just touch on a couple of things that you raised there. So you mentioned risk, you mentioned like multiplayer and stuff like that. So I'll try and unpack that the best I can. I think when it comes to like risk, [00:12:00] there has been this like at the start when OpenClaw kind of came about, it was never designed for enterprise use.
[00:12:06] Vincent Koc: It was never designed for anything but like a couple of developers tinkering on a side project, right? So obviously, it had inherent security risks, like your channels and messages weren't entirely safe. then everyone decided to kind of use us, our project and the hype as a little bit of a punching bag.
[00:12:20] Vincent Koc: Every security researcher, everyone's like, "Oh my God, look, did another security vulnerability," or something like that. And like, I think like 90% of it was complete slop. So we've-- if you go on the openclaw.ai website, there's a security page. It explains like all the vulnerabilities that were... The, the volume, where they're coming from, what they're responsible for, things like that.
[00:12:39] Vincent Koc: And, and, and like majority of that has been a big headache for us, and it's like this constant story that we have to keep telling people that like, I don't think this is any more unsafe or safer than another coding agent that's out there. Like we have insane volume of security researchers from some of the top research organizations and top tech companies that are working with us as [00:13:00] maintainers. yeah, we've done a lot of work. So we've actually been contributing to security, uh, around skills more publicly. So we've been like releasing data sets and releasing research that's helping the community, like how do we make agents safer and things like that as well. So that's the trust aspect. I think that's something that's like core to us and something that we deeply understand. But also on the flip side, of the missions of the OpenClaw Foundation is to make this thing accessible and universal and, how can I say, like yours, right? So I think having a solution that's open source and that you control and own, and you can use any model, you can use any channel provider, you own the data, is very important to us, and I think having that differentiation in market is gonna be very, very good for people, I think from a, from a... You know, I just wanna know that the agent that I built and the data that I've collected is like in my control, and I, I decide which, which model I wanna use. I decide where I wanna send this data. then when we kind of take that into an enterprise setting, going into things like [00:14:00] multiplayer One of the things I want to sort of backtrack and sort of explain a
[00:14:04] Andrew Zigler: Yeah.
[00:14:05] Vincent Koc: that when we were building OpenClaw ourselves internally, everything that we've sort of started to f- to feed into OpenClaw and build around OpenClaw was just because of our own internal frustration. talked about, you know, Steve Yegge's, uh, and Beads. If you don't know what that is, do have a look. It's like the early versions of looping and, and all that methodology and things like we've, we've spoken about. But we're essentially doing that. We were like kind of essentially like looping and like, you know, going full 100% on agentic engineering, but then quickly found out there was like issues.
[00:14:38] Vincent Koc: Like for example, how do we test this stuff? Okay, we don't have like sufficient CI testing infrastructure. What's out there? Well, nothing's kind of out there, so we're gonna have to build it. So we built this thing called Crab Box that can like use any of our, cloud providers and testing infrastructure.
[00:14:53] Vincent Koc: So we started building like these things around the frustrations. But then very recently what ended up happening is we kind of had OpenClaw at a certain [00:15:00] point But internally we were like unquote meat proxying the workaround. So we're like, you know, my agents are working on it, I'm passing it to you. Um, I started experimenting on what multiplayer could look like. Peter was definitely sort of experimenting, different people were working around this. One of the demos I did a little while back was like, uh, I could have a codec session, and it would work a bit like a reverse tunnel, and you would get like a web link or a QR code, and you could scan, and you can like re-remotely like take over my session. And there was this idea like, can I control into other people's sessions? Can my agents control other people's sessions on the team? And then we quickly realized, actually, no, this should live inside of OpenClaw. We should just all be able to like work on an agent in a multiplayer environment. So what we have now is itself is able to work in sort of this multiplayer environment, but what we've been doing is internally doing this.
[00:15:52] Vincent Koc: A lot of our workloads are running on a, on an OpenClaw server itself, and we're using that to build OpenClaw, and that [00:16:00] server or different OpenClaw servers, shall I say, are basically we can all jump in as a team and coordinate and take over other people's work. So if someone goes to sleep or there's an issue somewhere, someone goes away, we can go in.
[00:16:12] Vincent Koc: Or, or someone wants commentary, they can tag us and we can kinda jump in and give feedback. So what's happening here is like the way we manage work completely changes, is like the le-levels of ownership and trust and, and how we work as a team completely shift. This also means that like agents that are part of the team as well, like other claws, can also potentially co-participate and, and do this.
[00:16:31] Vincent Koc: So for example, they might move work from one box to another, or they might, you know, do certain things around that, like help us with certain things. But what we get-- what we've essentially done as a team is started building OpenClaw with OpenClaw itself. We've, we've essentially moved all our engineering effort so that the engineering work happens within OpenClaw itself rather than like a traditional coding harness, and this has kind of allowed us to really elevate our game.
[00:16:57] Vincent Koc: And this is why we kinda spent [00:17:00] two months or a month and a half for our last release did kinda nothing really. People were like, "What's going on here?" We just kinda felt like the experience could've been like really good, and we just wanted to make sure everything was good. So I think it was like something close to like 70 or 80% or 50%, like a large-largest proportion of our la-latest PRs and commits were in this one release.
[00:17:18] Vincent Koc: Um, something close to like 20,000, 30,000 PRs or s- uh,
[00:17:23] Vincent Koc: an, an, an obscene number
[00:17:25] Andrew Zigler: Yeah
[00:17:26] Vincent Koc: that sort of contributed to this. So... that's why we subsequently ended up calling it this sort of 2.0, but really the, the story was around multiplayer. The story was around enabling people and enterprises to be able to sort of work at this like higher
[00:17:37] Vincent Koc: level with agents.
[00:17:39] Andrew Zigler: Yeah.
[00:17:40] Vincent Koc: Yeah.
[00:17:41] Andrew Zigler: Using OpenClaw to build
[00:17:43] Andrew Zigler: OpenClaw makes sense as a building primitive, as something you're gonna pass back and forth. And what you describe is really fascinating because it does change the ownership narrative. Now the ownership becomes almost like this baton toss. It's almost like no one... Hot [00:18:00] potato, no one drop it.
[00:18:01] Andrew Zigler: Like, we're gonna pass it around in a circle and make it better and better and better. And that's a whole different experience than how we think about developing the code now, where I have my cold pool of the code, I have my work trees and forks I push back into thing. In, in this case, we're actually thinking about handing off the live hot interactivity of the artifacts and the things that are getting built and being responsible for molding that context together.
[00:18:29] Andrew Zigler: It goes back to, like, the conundrum that I struggle with of, like, how does my Gastown call your Gastown? Of what you're saying of, like, how do I hop into someone else's sessions and control something else? It fundamentally needs to get addressed at almost a different level.
[00:18:42] Vincent Koc: Essentially moved from Gas Town in my mind to like what I would call Gas City. I haven't read the Steve Yegge stuff in a while, so I'm not sure if he's actually r- talked about Gas City, but it's like we have Lobster-
[00:18:52] Vincent Koc: Lobsterville essentially, and in there there's little towns. Like you might have your own little neighborhood in there, but that's fine.
[00:18:56] Vincent Koc: But I can come visit your neighborhood, you can visit mine. We're in the same, [00:19:00] same,
[00:19:00] Vincent Koc: place, right? Um, so it's essentially for those that follow the law,
[00:19:04] Vincent Koc: yeah, that's kinda, that's kinda what it's going,
[00:19:06] Andrew Zigler: There, don't worry, you're in the house of Yegge, so
[00:19:08] Andrew Zigler: everyone everyone that's listening is probably pretty well familiar. Uh, we, we've covered this stuff pretty extensively here, and I completely agree. And, and, and that's what's really interesting to hear from that y'all too is u- using it, uh, to develop it itself.
[00:19:21] Andrew Zigler: I mean, it sounds like a, a virtuous way of continu- continuing to, um, develop the technology. But I, I'm curious as well about knowing, like in that feedback loop, it goes really quickly. I'm curious, what are strategies that you have for, uh, like I'm curious about the artifacts it leaves along the way.
[00:19:44] Andrew Zigler: You're passing this hot context that's doing things like shipping code and getting things out there hot and ready, but we as humans, we still need these artifacts and these ways of glimpsing in and, and having a shared understanding of the state. Uh, what strategies do you have for folks leveraging this kind of, uh, [00:20:00] hot potato agentic technology, I guess, this multiplayer experience to stay on the same page?
[00:20:05] Vincent Koc: There's a couple of things we've built over time that sort, sort of feed into this. So like the first one, this was prior to even us having like all the, this levels of automation and systems. Like we had at a point where we had, I don't know, 20,000 PRs were open on the repo. And we had, and we had rate limiting issues. Um, so one of the things we had is this thing called Git Crawl, which essentially is like a distributed network of everything that's current within our GitHub database, clustered, embedded, rinsed by LLMs. That's like a shared distributed understanding network and like lives on this, the Cloudflare worker, also works as a static sort of database. So our agents are aware of like all the work that's going on from that. We have something similar for our Discord community, so our agents are all aware of like what's being discussed inside of Discord at any given time, also between maintainers. So we have like [00:21:00] these like essentially like these tools or these like ways of collecting information and sharing information in an agentic manner that we've built over time and been able to share.
[00:21:10] Vincent Koc: Mind you, all
[00:21:10] Vincent Koc: this is open source, by the way, under MIT
[00:21:12] Andrew Zigler: Yeah
[00:21:13] Vincent Koc: welcome to go into our org repo. We have over 100 repositories, by the way, that solve different parts of this puzzle problem, so do take a look. And were kind of like lay- layering those up, and that's what's kind of enabled our agents to be aware of each other's work really, really well and be ch- and be kind of mindful of like what's current and up to date. And then we can kind of like layer that on with like this high agency, high throughput workload, which is like if I'm working on something and it's taking me a whole day, but someone else is already halfway through it and have solved another part of the puzzle, it might just take over my work, and that's kind of okay, and I'm, and I'm fine with that, right?
[00:21:48] Vincent Koc: Or vice versa. And kind of creating this like working open, high speed, you know, your work trees are on a shared infrastructure of change, changes that. I think it's less about moving the baton. [00:22:00] You still take ownership of your work, and you're still kind of making sure that that thing that you were trying to achieve was done or that feature or whatever it is.
[00:22:06] Vincent Koc: Like you might see other people's agents sort of come in and out of your work, or other people might come in and out of your work. But it kind of, it kind of increases the output speed and changes the, the way we operate quite a bit. from an artifact perspective, like everything is contained and in the traces.
[00:22:25] Vincent Koc: I still have like, I don't know, my last six months of traces still on my machines, and it's t- it's, it's kind of worrying. I do need to move it off because it's like getting to a point where I don't have the
[00:22:34] Vincent Koc: storage to keep up with it. But in a team setting, we might find that certain changes are breaking things.
[00:22:40] Vincent Koc: Well, okay, let's learn from that, or guess what? We have all the traces so the agents can identify back. can build memory off these. We can do some interesting sort of conceptual things to create like a shared understanding of what that looks like. Or if there's a new issue, can kind of get feedback.
[00:22:56] Vincent Koc: Like one of the things I'm toying with at, at the moment, which is gonna terrify people, [00:23:00] is that like we have a team box, sure, but what happens if like one of the threads can be made public? What happens if I can invite people from the outside to be able to come in and like leave comments and like give feedback or like look at something? So instead of me taking that from someone and then copy and paste it into that, like again, like this reduced the friction as much as possible. this is where I've given examples and working towards like, let's say I'm working on a fix for something that you've requested publicly on X. I can just give you a link and you can watch it happen in real time, and you can give me feedback as well like, "No, this is not what I meant," or, "No, this is exactly what I meant."
[00:23:33] Vincent Koc: And I think that's changing the way we, we think and way we operate, and, uh, I think it's more about reducing the friction for that information exchange and getting the information to and from those individuals like as quickly as possible where there's less people in between that a- as, as it needs to be. Like if it doesn't need to involve the humans and it just needs agent to agent in there, let's make that happen. Let's facilitate that if we know what the
[00:23:58] Vincent Koc: information needs to be exchanged. [00:24:00]
[00:24:00] Andrew Zigler: Yes, big plus one, too. I don't want us all just end up being the meat proxy, as you said earlier. I think this is actually what has driven so many to experiment with the technology, because they find themselves in this copy-paste pivot position where they want, someone wants access to whatever agentic riches they're perceived to be sitting on top of.
[00:24:18] Andrew Zigler: And so I, I, I think that that's, like, a, a really big thing to call out. I also personally resonate with it. It's why I've used OpenClaw inside of, like, o- my own kinds of projects to make it easier for folks to just engage and drop in themselves. Uh, and it also, actually, what I've learned is it sharpens the ask, because when folks come to a collaborative space, there is a sense of shared responsibility and, like, we're managing now this experience, this project, this outcome together.
[00:24:48] Andrew Zigler: Um, and that allows somebody to sharpen what they really want, be precise. There becomes a skill involved also in contributing, and new expectations for, like, how you bring your ideas to the group. [00:25:00] And that's what I think is really, really interesting, just, like, when you talk about in general a multiplayer AI experience.
[00:25:07] Andrew Zigler: Like, a little background on me is, like, even before I worked, uh, at LinearB and, and my first, uh, one of my first jobs in tech was at Mattermost. It's an open source, uh, communication company, and it's like self-hosted chat basically. And a big part of what I was experimenting there when ChatGPT first came out was what does it look like when you put an LLM in a self-hosted chat channel where everyone is there together and you actually all create a shared context?
[00:25:34] Andrew Zigler: It's a really fascinating experience of just, like, you- you then become responsible for, uh, crafting that kind of experience together. It makes also everyone a builder, um, and I know that you've encountered this, of where you're probably bumping into people who come from a totally untraditional builder backgrounds who are using the technology that you've built to do some really incredible things.
[00:25:56] Andrew Zigler: So what do you think about the power of domain [00:26:00] expertise when applied to a primitive like OpenClaw? What opportunities do you see for the world still to really get on top of this technology and, and, and find uses for it?
[00:26:12] Vincent Koc: Can you repeat the question again? Sorry
[00:26:15] Andrew Zigler: Yeah. So what opportunities do you see, like in a broader world setting for folks to use OpenClaw as a collaborative way to share, and form, and sharpen, and deliver ideas? Because that's, I think, the power of it is that you can use it so well to even develop itself. You can truly use it to sharpen and develop any idea.
[00:26:36] Andrew Zigler: Um, this becomes a really powerful thing for folks to leverage even just in a collaborative way. So what, what opportunities do you see just like for the broader world? Because I feel like in the tech world we get really excited about how it's applied for our immediate use cases, but I just do also think that it has a br- a lot of broad,
[00:26:52] Vincent Koc: I
[00:26:53] Andrew Zigler: implications as well.
[00:26:54] Vincent Koc: for artists. Like even if you think about like a, a, a, an association or a village [00:27:00] community or whatever it may be, multiple people
[00:27:02] Vincent Koc: need to share information. But if something has that
[00:27:04] Andrew Zigler: Yeah
[00:27:04] Vincent Koc: or that shared understanding of a given space or a problem, it can definitely help propel us forward. Um, I keep going back to this concept of Wiki, and there's a, there's a really good book that kind of talks about Wikinomics and how we solve problems with like the way of Wiki, which is essentially this whole idea that, you know, if we have a shared distributed understanding of a problem and we open up that information, we can actually solve problems a lot faster and a lot better with individuals. And this sort of concept sort of applies really well because it's like actually we're kind of elevating that process with
[00:27:32] Vincent Koc: agents and AI, and we're kind of
[00:27:35] Andrew Zigler: Yeah
[00:27:35] Vincent Koc: sort of shared consciousness or that shared understanding of a given, uh, issue to sort of come to life. So really it... the use cases are like as large as it can be, but it's more a case of like how do you bring these individuals on board? what I'm seeing right now is like individuals building Claw-like technology, but like being really opinionated about the user interface and like targeting a specific demographic of user
[00:27:58] Andrew Zigler: Exactly
[00:27:58] Vincent Koc: becomes like [00:28:00] the, that, that version of what that looks like. Um, I definitely see a world where individuals are gonna have their own AI personal agents. How does that look and shape is, is still to come, but you still have to think like a huge proportion of the, the population globally has yet to sort of interact with AI on like a deep personal level like we're talking about. So we're still talking about like at least like I, I, I don't, don't quote me on the numbers, but I would assume like huge percentage or a great magnitude of, of the world's population is still yet to sort of experience what we're
[00:28:30] Vincent Koc: experiencing with artificial
[00:28:31] Vincent Koc: intelligence.
[00:28:32] Andrew Zigler: Exactly. Exactly. Yeah. I think that's just, like, the real opportunity, and I think frankly they're gonna walk through the cl- the doors of OpenClaw first, and I think that's going to be, uh, a really interesting thing, thing to see. I'm really happy that you brought up wiki, because that was really where my mind was racing to next, is the, the wiki-ification of it all.
[00:28:54] Andrew Zigler: I love a good wiki. I love organizing information. I'm a classics [00:29:00] major by training, so there's nothing I love more than reading a bunch of stuff, labeling it, coming up with a hypothesis, and having a bunch of little sticky notes out of a super old book. So, like, I'm really on board with the idea of, of having the wiki and using that as a collaborative, uh, format.
[00:29:15] Andrew Zigler: And I think that when you combine it with an agentic world, when agents crawl across, create this wiki, make it up to date, connections appear where they weren't, it becomes a really amazing representation of just, like, the vector space even that they live in, in a, in a way that touches really closely on what we can read and navigate as humans.
[00:29:35] Andrew Zigler: Um, like I, I myself, I have like a really fastidious journaling process. I t- use a, I do a dot journal that I've kept at Obsidian long before that was en vogue with, to do with agents. I did this like five years ago, and I've just done it every day. And, uh, I use this as like a note-taking system in combination with what's called a zettelkasten, which is like a note-taking system that's super flat.
[00:29:58] Andrew Zigler: Um, all of my notes are [00:30:00] interconnected with each other.
[00:30:01] Vincent Koc: I, I
[00:30:01] Andrew Zigler: they're not sorted by any kind of topic.
[00:30:05] Andrew Zigler: Yeah
[00:30:05] Andrew Zigler: so it's like you're familiar with that world. So I already think in this massive sea of slightly connected ideas and arcs that have accum- accumulated over time, pinned by this whole timeline of documents.
[00:30:17] Andrew Zigler: And so honestly, this becomes almost like a vectorized representation of how I think over time. Using that as my mindset of how I come in and I use a, an, um, AI to collaborate and build as a, as a GTM engineer with my team doing this, everything behind the scenes here at Dev Interrupted, which is highly agentic and runs with a whole bunch of pipes, uh, behind the scene, right?
[00:30:39] Andrew Zigler: Is, like, applying, uh, applying that kind of organization method to get other collaborators involved. Uh, it creates this semantic language. That's the power of the wiki, um, is it allows you and, and agents and workflows to all express in a shared understanding this web of, of knowledge. And OpenClaw, [00:31:00] I think, is really what pierces into that and makes it then something that people are chatting with collaboratively, which is a, a really e- exciting thing to see
[00:31:14] Andrew Zigler: You say yourself have a, a note-taking kind of system like this. You think in notes like that as well?
[00:31:21] Vincent Koc: geez, like a long, long, long time
[00:31:24] Vincent Koc: just text files, like really
[00:31:26] Andrew Zigler: Yeah
[00:31:26] Vincent Koc: very similar to what you were saying.
[00:31:28] Vincent Koc: Um, so this
[00:31:30] Andrew Zigler: Exactly.
[00:31:31] Vincent Koc: like
[00:31:32] Andrew Zigler: I just think what's the most f- what's the most future compatible way of holding my ideas over time? I just write them down in markdown files, and I have for a very
[00:31:41] Vincent Koc: that's
[00:31:41] Andrew Zigler: many years
[00:31:42] Vincent Koc: it's obviously easy because you have the history of the markdown files. But if we take something like a memory, uh, in a markdown or something like that, what's missing is that like, how did this, come about? Like, what's the intent behind it? What's the thinking? like [00:32:00] that's constantly missing. I think we need to sort of think about like, how do we capture that? How do we capture the why I did this thing? And a lot of that
[00:32:08] Vincent Koc: is based on my personal experience of that, that day, that week, that month, that project, the individuals I'm working with. I need to capture a lot of information to try and like make sense of that. But I would say like cap- that's the thing that we need to sort of capture is our essence of why. Because I think as soon as machines can understand why we're trying to do something, um, or there's like multiple individuals and there's a lot of different whys, I think we're then able to sort of like calibrate a little bit better, um, or at least achieve a better outcome.
[00:32:36] Vincent Koc: So that's an, an interesting space that I'm sort of messing around with. Can I infer the why? Um, I had another side project. I did a, this sort of like example. It's called Tideclaw. It re- it's on my personal GitHub org. And, um, this basically looks through like all your messages and your Codex logs and tries to infer like your personality and like why you do certain things, and it tries to [00:33:00] capture that in sort of markdown outputs. And then you can use that as like steering or like instructions to agents to say, "Okay, this is my preferences for like when I code, when I b- work on a project, when I communicate with other people and, and vice versa." I think trying to understand intent becomes quite interesting as well for like the preference aspect and like how can agents infer that for you without even having to ask you using other types of information that you might already have on, say, your machine or your messages and things like that.
[00:33:30] Andrew Zigler: Yeah. You know, you bring up the question of why. This becomes a really important thing for collaborators to know, and that's where does why live? A why is, can be attached to a lot of different decisions along the way, all which need their own kind of why. And
[00:33:47] Andrew Zigler: I, I, I think, like, an important one is, like, you, you said, uh, maybe you make a code decision or something is implemented in this way, and you wanna know why we made that decision.
[00:33:57] Andrew Zigler: Maybe it was something that was part of the [00:34:00] coding session. Maybe it was part of the, the planning session. Maybe the why lives in when the code was created, but maybe the why was back w- uh, wherever the specs live. Maybe the why is actually a larger question than the specs even see. Um, so in this, like, agentic experience where people are all kind of coming in from different, uh, vantages, where do you think why lives, and
[00:34:22] Vincent Koc: our
[00:34:23] Andrew Zigler: how do you think that that, uh, survives?
[00:34:24] Vincent Koc: of all of the messages and conversations we're having that's re- recorded and transcribed and into these, uh, databases that I'm talking about, and then couple that with the, team sessions that have access to those same information. So then as we start to steer them, it, it has this, not in a clearly set defined why, but it has the ability to go and search out and seek information that will tell it, you know, how this, this thing originated, I think is interesting for the way that we've set up our workflow, and I think that's why we're getting some really interesting results internally
[00:34:56] Andrew Zigler: Absolutely. So you're able to really have a loop that's able to then pin [00:35:00] conversations and intent and alignment into the actual work that's getting done. That's pretty co- that's pretty cool. Um, I, I've done some experimenting around trying to do some of this as a solo person using, like, beads,
[00:35:12] Vincent Koc: Mm-hmm.
[00:35:12] Andrew Zigler: which you mentioned earlier, and track- doing task tracking through beads and having that as, as a clear record of why for certain things.
[00:35:19] Andrew Zigler: And
[00:35:19] Andrew Zigler: it's interesting when in a multiplayer way of that, of being able to be something that is, like, encoded and shared with other people, and really it speaks to the, the real... I think the, the biggest transformation, especially for folks that are maybe, like, on a more, like, a go-to-market side as well, a b- a big transformation for folks is that now all of the work is just around alignment.
[00:35:44] Andrew Zigler: Before it was like, "Oh, I'd have to need this timeline to get this asset done, to get this across the finish line, to do these blog posts, and get all these social things done, and then these designs. Oh, there's all these timelines." And, and everything is, it has this window. But since the compression of that window is so [00:36:00] small now because so much of that can be run in parallel, and actually so much of that can be done on the instant that we know what we want, now all of the work is in the discussing, in the aligning, uh, in making the decisions as a group.
[00:36:13] Andrew Zigler: Which is why then speaking to what you're saying, it's important that that happens alongside the agents, so they can understand the why. They can pack it into mechanical checks and balances that they can leverage. Um, otherwise, you know, it's just all gonna escape as, like, vapor, and it's not gonna be useful.
[00:36:32] Andrew Zigler: It's kind of like you're not then truly capturing the back pressure. Like, when we had Geoffrey Huntley on the show, uh, back when the Ralph loop was transforming the world, you know, uh, what simpler times that was. Uh, it, it r- really was exciting to really, uh, hear from how he used that to capture the back pressure, and this is talking about, like, his famous experiments around doing this were back when, uh, like, Claude was particularly squirrely, not super good at using tools.[00:37:00]
[00:37:00] Andrew Zigler: Yeah, it was, uh... It was quite devious actually going through the cursed repo and seeing all of the scaffolding that he had done to make, uh, that language come into being,
[00:37:10] Vincent Koc: and I
[00:37:10] Andrew Zigler: is truly is a feat
[00:37:12] Vincent Koc: how, why "This even works?"
[00:37:14] Andrew Zigler: It's, it...
[00:37:17] Andrew Zigler: Uh, why does it work? Uh, I was so inspired by it,
[00:37:19] Vincent Koc: it works,
[00:37:20] Andrew Zigler: made, like, a, a DSL server.
[00:37:23] Andrew Zigler: I-
[00:37:23] Andrew Zigler: it's, it's, it's devious.
[00:37:24] Andrew Zigler: It was devious
[00:37:26] Vincent Koc: out of it. There was a lot of this looping. I think there was like, um, other thing, I think Root Code was another one, which was like a harness that kind of had this built into it, which would like try to build a loop mechanism into it. But then I think at one point the models got good, and it was around the time that like really OpenClaw took off, and I think that's kind of part of the, the puzzle where, you know, loop- the models were good enough that actually looping actually kind of started making sense. I think right
[00:37:49] Vincent Koc: now it kind of, it's like a no-brainer. It's like I just couldn't
[00:37:53] Vincent Koc: see any reason why you wouldn't, you would not do
[00:37:55] Andrew Zigler: Of course
[00:37:56] Vincent Koc: Um,
[00:37:58] Andrew Zigler: Exactly. [00:38:00] But that was the first taste of it. That was the first taste of all of us being like, "Oh, actually this loop thing is starting to work." Even to your point, like Anthropic even fully leaned in. They had an official loop plugin. They had the Claude or the, like the Ralph plugin, where and all it was, that was loop, right?
[00:38:14] Andrew Zigler: And now that's an expectation. Now we call that the long running, the, the, the task, uh, running, you know, agents. Now it's a whole classification of benchmarks. We don't even think about it that way. And so it's, it's fascinating just to see how f- how far, um, everything has developed this year, OpenClaw, uh, included.
[00:38:33] Andrew Zigler: I, I'm curious, like, uh, you know, Vincent, you have a pretty impressive background and you, you've been at a lot of places and you've seen a lot of things, but most importantly, you've also been, uh, in and out of the classroom and thinking about how do I bring the, the understandings that I have about the data world, uh, into, uh, the, like the next opportunity of learners.
[00:38:54] Andrew Zigler: And that actually goes back to, I think, so many of the deep-rooted, uh, elements of what is going [00:39:00] to make OpenClaw continue to just be more relevant is that, uh, like, uh, it's very well adopted among newer and, and learning engineers and people that are learning how to build. Uh, I think that really leans into your instinct as a teacher, and so w- how do you bring that into building and, and running like,
[00:39:17] Vincent Koc: I
[00:39:18] Andrew Zigler: OpenClaw and thinking about it?
[00:39:19] Vincent Koc: a lot of teaching actually funnily enough is like less about me telling people what they should or shouldn't do, but it's actually like listening are not understanding and then like how do I fill that void and how do I take them on the journey? uh, for everyone's context, I teach at universities, um, mostly postgraduate around, uh, agentic AI and artificial intelligence, everything from deep learning to like agents.
[00:39:41] Vincent Koc: I remember even I used to have classes in the early days of like LLMs, like, "Oh my God, let's fine-tune a small model," or like, "Let's, let's build a rack." And it was like fascinating, right? Now this is just like everyone kind of understands how that stuff works even if you don't need to like go into the code, but you, you conceptually can understand it. But I think it's just a, [00:40:00] a misunderstanding or like getting a level head, a level footing in terms of like what's happening. So I think a lot of teaching is about sort of taking people on the journey and like helping them see where the gaps are, but you kind of need to be able to listen and understand that.
[00:40:13] Vincent Koc: So I think the key for myself and the development work that I do in, in OpenCore is really about like really dialing into like what are people like not understanding here? Like where are people getting stuck? Where is it like not making sense? I think sometimes it's about crafting a narrative, like let's call it multiplayer, let's call it 2.0, and then suddenly people are like, "Oh wow, this thing's changed.
[00:40:31] Vincent Koc: It's different." or it's like more about like how do we change the experience, and how do I put myself in the shoes of the users? And I think this is just like a constant battle of like trying to adapt. sometimes what you're gonna find out is gonna be things that you don't wanna hear. It's gonna be bad and ugly, or it's gonna be things that kind of challenge your thinking about what the, what you think is the truth. But I think sometimes it's, it's kind of healthy and refreshing. Easy for us as an [00:41:00] organization or as a team because we are quite small and we're moving really fast, that we have this sort of license to be able to adapt and break things and just kind of... Break things in the sense of not break it for users, but like, change the way we do things But, um, I can see how that might be more challenging in a, in a much larger organization that's been around for quite some time and it's used to a certain pace or it's used to working in a certain way, and then you kind of come along and go, "Let's just use agents to do everything.
[00:41:24] Vincent Koc: Let's, let's slash loop on this and go Ralph loop and beads and everything else." So I that's where I don't have the answer. It's like, how do you an organization mature? And in the past when I've done some like this ugly word called digital transformation, which just basically means change, um, has been like, find a team that is like so willing to like dog food this thing and start small, find a project, start small, and just like grow from there.
[00:41:49] Vincent Koc: Don't try to like convince your whole entire company to like start looping everything. Don't convince like everyone to like embrace AI. Like just start one place, fix the frustration, solve [00:42:00] the problem, show the ability and the capability of what it can do, and then just go from there. I think this also applies to people trying to do AI in their personal lives with personal agents.
[00:42:07] Vincent Koc: Like they're throwing everything under the sun at this thing. But actually like if you can just get it to like summarize your emails every morning, that's a win, right? And then like just layer things on bit by bit. Don't try to like solve everything at once. and I think that's where you're gonna start getting some value out of a lot of this technology.
[00:42:23] Vincent Koc: That, that's my 2 cents on this, on, on, as a, as an answer to your question.
[00:42:30] Andrew Zigler: No, it's great advice. Just starting with realistic expectations, iterating quickly. Uh, it's really like what you said, doing small passes to add things one bit at a time. And that also then means, you know, making sure that you understand the things that you're adding and that it's making sure that it is aligned with what you want, and
[00:42:49] Andrew Zigler: stopping, you know, when, and, and realigning it when you need to.
[00:42:53] Andrew Zigler: It's kind of like the, the real craft of it is, uh, I've, I've alluded it recently as like you put it on a potter's wheel and then you're [00:43:00] shaping it, and that comes down to how often are you running the workflow? How often are you running the, the, the, the ... How often are you encountering this agent or using this skill?
[00:43:08] Andrew Zigler: That's what's gonna be your interval for, like, how, how much can you shape it? Um, and it's about, you know, bringing the right amount, uh, each time. Not too much or you're just going to make, uh, a little bit of a mess. And I, it's been a really amazing, Vincent, to get inside your head about how you think about, uh, building OpenClaw, how you think others should think about collaborating via OpenClaw.
[00:43:31] Andrew Zigler: I think there's a lot of opportunities that you've showed us today just for the teams to get on the same page, uh, to implement workflows that allow why to live alongside the work that gets delivered, which I think is the holy grail for a lot of folks trying to be agentically enabled. Um, and I know you're constantly, uh, pushing out a lot of new, uh, like, research and blogs and, and, and discussions around, uh, what you're building.
[00:43:55] Andrew Zigler: And so do you have any, uh, you know, parting last bits for our, uh, [00:44:00] listeners for where they can go to, to stay tuned for the latest for you or, or, or keep in mind for some upcoming things that you're excited about?
[00:44:06] Vincent Koc: of keep an eye out on announcements. It could be on X, on, on Reddit, or just even on, on our website as well. But yeah, we're definitely gonna start punching out some really interesting work, over the next, next few weeks and next few months that I think is gonna get people really, really excited. And if you haven't used OpenClaw in a while and used it before and it wasn't the best experience, I'd say give it, give it another shot. It's a, a very different experience. I think we're moving away from messaging channels. We're moving away to this like native web app sort of ecosystem where it's like, wait, why, why do I need to be inside a Telegram?
[00:44:39] Vincent Koc: I think it's good when I'm out, out drinking coffee with a mate, and I just wanna be able to send the agent something, but otherwise I wanna be able to see my work in one place in this sort of like thing. So we're kind of shifting away from terminals and other sort of things into like more purpose-built, um, user interfaces. One last thing I wanted to add, and you know, we haven't had the chance to get into a demo given the time that we [00:45:00] had, and maybe we can attach some screenshots or something,
[00:45:02] Vincent Koc: um, to this later, but is, um, dashboard for example, the, the, the, the OpenClaw web sort of interface that we're working from has the ability, like you can have plugins that can change the entire user ex- user, user experience, but also your OpenClaw can do that as well.
[00:45:19] Vincent Koc: So we've got one internally, for example, creates a newspaper for us every day of like what's changed on OpenClaw or things like that. So you can basically like reimagine what your interface should look and feel like, and it will build that for you. So there's people on the team who've been building some really interesting things around like me like my day that's coming up and like use visual blocks and, and things like that.
[00:45:39] Vincent Koc: So we're kind of thinking about like this whole smart home sort of like devices that you have kind of experience. Um, but the AI is sort of driving that for you based on what you want. So, and then we're thinking about how do we then bring this into like mobile devices and things like that, so your agent is around you and able to kind of share that information with, with, you and interact with you as [00:46:00] well.
[00:46:00] Vincent Koc: So I think gonna be some interesting, interesting developments, but honestly at the same time, like who knows what's gonna happen in two weeks or three weeks' time in, in the AI space. There's always some interesting developments and for our team especially, it's like how quickly can we react and implement that and sort of be on the edge with, with something like OpenClaw?
[00:46:19] Andrew Zigler: Yeah, that sounds really exciting. I, even what you described is almost like the building blocks of, like, a personal factory or a bigger factory. This idea of being able to arrange the parts that you need, uh, uh, i- and, and really thinking about, uh, owning the, the interface by which you interact with your data and your context of what matters to you, I think that becomes the new fundamental builder responsibility of everyone who really leans
[00:46:43] Vincent Koc: Yeah
[00:46:44] Andrew Zigler: in the moment.
[00:46:45] Andrew Zigler: And it's been really exciting to see what that's gonna look like. And to our listeners, you know, we're gonna include links in our show notes where you can go and check out all of this technology, including the latest features available in OpenClaw 2.0, where you can dive into more and s- and get a closer look at exactly what [00:47:00] Vincent is talking about.
[00:47:01] Andrew Zigler: And, uh, to those as, as well that were intrigued by our conversation, we'll include links to the foundation, so that way you can get reacquainted with OpenClaw if maybe it's been a hot minute. And if you listened to the conversation all the way to this point, then you clearly had a blast, and you're a total either Steve Yegge nerd or an OpenClaw nerd or both.
[00:47:19] Andrew Zigler: So wherever you're listening or watching this, you should give us a thumbs up or a subscribe. And come find Vincent and I on socials. Uh, just come bother us about anything that we said today. Take a side. We don't care. Just pick a side. Join the conversation. Uh, and we wanna hear from you. And thanks again for listening to Dev Interrupted.
[00:47:37] Andrew Zigler: And Vincent, thanks for joining us. It was a joy chatting with you today
[00:47:41] Vincent Koc: on the show
