0:07 Good morning, good evening, wherever you are, and welcome back to the Enterprise AI Show. 0:12 This is your host, Aaron. 0:13 And today, Brandon sits down with Adam Jacob, CEO and co-founder of SwampClub and the creator of Chef, to talk about why AI should change the way we build software, not just speed up the way we already do it. 0:30 Adam explains what it really means to build the machine that builds the machine, and why his engineers now average 35 to 45 pull requests a day, and why he thinks nobody is actually behind. 0:46 He also shares how SwampClub cut the token cost of one code review workflow by about 10 times. 0:54 And all of that is coming up right after this quick break. 1:02 Today's show is sponsored by Nord Security. 1:04 Your employees are using AI every day, but do you know which tools they're using, what data they're sharing, or where their AI agents are connecting? 1:11 You can't govern what you can't see. 1:14 Introducing NordLayer AI Scout, the visibility layer for enterprise AI. 1:19 AI Scout gives you visibility into AI activity across web and desktop apps, detects sensitive data exposure, helps control AI agent connections, and creates an audit trail for governance so you can protect your data without blindly blocking innovation. 1:32 See what your AI is doing, then decide what happens next. 1:36 Connect with Nord Security to learn more about NordLayer AI Scout. 1:41 Visit nordlayer.com/ai-scout/enterpriseai and get started today 1:53 All right, joining me today is Adam Jacob, the co-founder of Swamp Club, also a returning champion, as I like to say. 1:59 He's been on the show before. 2:00 So welcome back, Adam 2:02 Hey, thank you. 2:02 I'm pleased to be a returning champion 2:05 Well, listen, as everything that happens with technology, the world has changed a lot since last time you were on, and we're gonna get into a, a bunch of stuff. 2:12 Some some things you're doing now and a good presentation you got. 2:16 But just for everybody that maybe doesn't know you, give us like the the 90-second tour of who you are and what you've done. 2:23 Yeah. 2:24 So right now I'm the CEO of a company called SwampClub. 2:28 SwampClub basically helps you build sys- AI systems out of deterministic behaviors. 2:34 So basically you use frontier intelligence to write these deterministic behaviors, and then you put those things together into workflows, and then you call out to frontier intelligence. 2:43 So it saves you time and money and focus, and actually lets you build systems you can understand and refactor. 2:49 That's the short version of SwampClub. 2:51 In a previous life I… Most people, if you know me for anything, you know me because I wrote Chef originally and I was the CTO of Chef. 3:01 And a lot, did a lot of work in the enterprise helping people understand what DevOps was what, used to be called digital transformation moving to the cloud, all of those things were things that I was present at the creation for and helped push into the universe. 3:18 And then, and in between I had a company called System Initiative that didn't quite work out, that sort of some of the people molded into SwampClub. 3:25 None of the technology did, but the, the people did. 3:27 Some of the people did. 3:29 And yeah, so entrepreneur, venture capital, open source, uh, and infrastructure nerdy. 3:38 Yeah, it turns out I just like the whole, I like the whole part of the game. 3:40 I like all of it. 3:42 I like the whole thing it turns out 3:44 we're gonna get into Swamp Hub 'cause I wanna get into the details of that. 3:47 But also, you didn't say it there, but I think you're also a very good speaker, a passionate speaker That I've seen sp- I've seen you speak several times, and more importantly, a lot of your videos are on YouTube. 3:56 So the one that caught my eye, and I, was just thinking about as, as I wanted to invite you on, is it's titled here "AI is the Future of Software Development." And then I think it says, "Every Objection You Have Will Be Overcome." And I just, I love it. 4:11 And the reason I wanna have that on is I think there's two things I've noticed, and I wanted to get your take, was like, one, I think there are people that don't know anything about AI, but those people aren't listening to this podcast. 4:21 I think there are a lot of people listening to this podcast that use AI, but they haven't, if you will, really, made it an intrinsic part of how they do things. 4:33 And I just, this, this presentation you had, I feel like, I would almost call it like a call to action. 4:38 It's almost like you think you're doing AI, but maybe you aren't. 4:44 And so maybe we'll start with a simple question, 'cause that's me editorializing a little bit. 4:48 What drove you to even create a presentation like this? 4:52 Yeah. 4:53 That presentation was that was when we first showed people Swamp. 4:57 So that was like the, literally the first day that there was an existing, a thing called Swamp. 5:02 And it was born out of my experience giving building system initiative. 5:10 That talk you're referencing was at a place called Configuration Management Camp in Ghent, Belgium, which is like my favorite conference in the world. 5:18 It's like at this community college in Belgium filled with just the most hardcore, like, configuration management nerds on Earth. 5:26 And I feel like it's like I'm there amongst my people. 5:30 I would say that talk was like moderately well-received. 5:34 There was like a, there was like a, there was like a, a kernel of people there that loved it, and then there were people, certainly people there who were like, Can I swear on this podcast? 5:42 I actually don't know. 5:43 They were like,"F this guy." You They were just like,"I don't know about that." I think for me, what I saw that talk I gave in February of 2026 so almost a year ago, Was that the, rate at which the models themselves were improving and the quality of the code that they wrote. 6:05 Everybody's first experience with AI was that you used AI inside your existing workflow. 6:11 You're a software developer, now you have AI. 6:13 Rather than write the code, you prompted it. 6:15 You looked at what it did, and mostly you were like, "Eh." and then every once in a while there was this glimmer of hope where you were like,"Oh, that was actually pretty cool." And as the models have gotten better, it's been harder and harder to ignore that those moments are constant. 6:31 It's, doing a pretty good job most of the time without a whole lot of hand-holding or a whole lot of, behavior. 6:37 But you mostly slotted it in, like I said, into your existing workflow. 6:41 You opened your editor, you were working on code, you understood it, you were thinking through how that stuff worked. 6:46 What became clear to me a little earlier than February, but still just roughly in that time period, was that rate of improvement and the fact that you could now build a machine whose job was to build the machine. 7:01 If you think about building the software that you want, the way we've always built that is by writing the software. 7:06 We sat down and we did the work. 7:09 What started to happen is you started to see this shape where it was like, actually, the software can adapt to your needs, and it can adapt to what you need it to do in a way that wasn't true two years ago. 7:22 And once you see that the software can be adaptive in that way, what then happens is you start to think about how does that impact my ability to build software, build products, ship things to customers? 7:36 And you start to build this machine whose job is to produce the outco- the outcome that you're looking for, instead of thinking so much about how do I build the outcome I'm looking for directly. 7:46 And it's that's a big head trip. 7:48 It's very strange. 7:49 And what I was trying to say to my friends at that configuration management camp and, folks who do it I'm an operations guy in my heart. 7:56 I still think of myself as a systems administrator, even though I've been a CEO now almost as long as I was a professional systems administrator, probably. 8:05 I hadn't actually thought about that till I said it just now, but it's probably true. 8:09 Or I-- certainly if you mash up being CTO and CEO, it's probably longer than I was a professional systems administrator by title. 8:17 But operations always lags software development. 8:23 And so the trends that happen in software development become the trend in operations with a twist historically, always. 8:29 And so this room of my peers, they were also in-- were tending to be the most AI skeptical and the most the most filled with doubt that, it could ever work for them or that it could impact their lives. 8:44 And I just, I really felt like they needed to wake up to the fact that this was a moment of great exploration. 8:52 It's a moment of great possibility, and we either jump on that possibility and we define what that world looks like for ourselves and for each other and have that conversation, and it'll be messy and gross and uncomfortable, or a significant number of people in that room who I really love, like I've been seeing every year at this tiny community college in Belgium for a, a decade more, like we're gonna have a very tough time in their career and in their personal lives because we love this work and we have to figure it out. 9:27 And the best people in the world I can imagine to figure it out together are the ones who were in that room when I gave that talk, and that's what inspired me to give it. 9:35 I like it. 9:35 I like, and I think you said something in there that I think is really important. 9:38 It said something like it's this mind shift from you trying to build to a business outcome. 9:44 But I think you said something like it's the machine to build the machine to build the, the thing that's ultimately outcome. 9:49 And I think that's an area that we should explore at least a little bit more, 'cause I think one of the things you say in your talk is I feel like it's very inclusive. 9:55 It's I think it even ends like,"Let's figure this out together. 9:58 Let's all get in the…" and I think you also make the point of a lot of people feel like maybe they missed out or they're behind, and I think you, you go on to make the point of like You're not behind. 10:07 If you're listening, especially if you listen to this podcast, you're definitely not behind. 10:10 There's no chance you are. 10:11 But I think one of the I'm gonna give a very simple example and I want you to maybe help us understand like how we can all adopt this idea. 10:19 And so what I see sometimes with s- let's say a very simple problem would be something like you have a list of like email address. 10:26 You have some random list of text. 10:28 Maybe it has email addresses, maybe it doesn't, and you wanna get all the email addresses to invite everyone to a party. 10:34 So a way sometimes people start, and maybe the way we all start it, is like first you go into maybe your Claude code or something, you say"Write me a script that will parse out all the email addresses," right? 10:43 That's how you… That's like step one. 10:44 And then you're like, you use it for a while and you're like what if I just gave it all? 10:49 I'll just give it all the text to be like,'Get all the email addresses for this thing.'" And then sure enough it does it and gives you all the email addresses. 10:54 And then you're like, and then you write the email and just… And then you're like, "You know what? Why don't I just have it like draft the email for me," right? 11:02 And then it's okay. 11:03 And then you get to a point of "Why don't I just have the thing like actually draft, send, and compose the whole email for me," right? 11:10 And then you get to the point of"Maybe I'll have this thing like organize the party for me," right? 11:13 And it's like you're just, like you have to walk up this, like this ladder, right? 11:18 And I think a lot of people go through something similar, right? 11:20 You Mentioned that before. 11:21 It's I'm gonna write one function, I'm gonna write a whole, like a whole library, and then I'm gonna write the entire thing. 11:27 So like what did you do or what do you-- Like what does someone need to do to really take that idea you said build it, that the machine to build the machine, to unlock that, that thinking process so 11:42 Yeah 11:42 if you will I don't know how to say it, in the swamp with us, they're they're they're trying to get there, if you will. 11:49 So there's an idea floating around called a software factory you might have heard of, which basically you can think of a software factory as like taking all the parts of your SDLC from how the product gets defined and how you think about the work and then how you would build it, and then how you would know that it is good. 12:08 How do you test it? 12:08 How do you verify it? 12:09 How do you ship it to production? 12:11 How do you verify all of that? 12:12 And putting all of that into essentially an agentic loop, right? 12:18 And that term, that idea of a software factory, sometimes there are lights on, which means humans are interacting with that factory at various points. 12:27 Sometimes there are lights off, which means the work comes in and it just goes all the way through to the end. 12:33 The way to start thinking about building this thing is that, A, you know what you want to build. 12:39 So you, the software developer, you, the product person, you, whoever, like you have an outcome that you want to see in the world, right? 12:47 That's, actually the source of your own creativity. 12:49 That's you saying, "Hey, I, I want to send this email and start a party." You know what it is you're looking to do. 12:54 How do you know what are the requirements of that being good? 13:00 And there's a bunch of them. 13:01 There are things like software architecture, right? 13:04 How maybe in your example, you don't care too much about software architecture, but in a big product you do. 13:09 You care about how it was built. 13:10 You care what the terminology is. 13:12 You care how the pieces fit together. 13:14 You also care about other things like maybe accessibility or observability, or you care about the various patterns that are being expressed in the language. 13:23 Whatever those things are, you start to encode those things as part of the agent's process by which it writes the code to produce the outcome you're looking for. 13:33 You start to see things like black box testing which is basically writing, having outside in tests that have no knowledge of the internals of the system, but all they do is prove that the product works the way you wanted it to work, so that later you can refactor it and know you didn't regress even if you don't know what all the pieces of code are in the middle. 13:51 And so the way you walk yourself up to that is you start thinking about it like an engineer and you start looking at the output of what's happening in the agent. 14:02 And when it's wrong, you start thinking to yourself: How would I build a system that makes this more correct? 14:08 And the, the fun part of that is that process is always unique to the thing that you're building, right? 14:16 So like what makes a great email party planner is unique to what Brandon thinks a great email party planner would be because you have a point of view about what that thing is. 14:27 And so you're encoding both that point of view, why, what makes it great, why is it good why, does this party planner work the way I want it to work? 14:35 You're also including the engineering principles that make that software good, that make it supple, that make it so you can change it according to your requirements You're also talking about how to, prove that it works, and all of that structure is the thing that then allows you to work on this machine that builds the machine. 14:55 It's the structure that builds the software. 14:56 And so when the software's wrong our instinct forever has been, as engineers, to fix it. 15:03 We dig in and we do this. 15:07 And now we're-- Now your instinct has to flip, and it has to be what went wrong s- structurally in the way that I build this product so that I can tweak that so that in the future that doesn't happen anymore. 15:21 So a great example there is if you've ever built a system that you wanted to have be highly accessible, so you want it to be like WCAG AA compliant, for example, to accessibility guidelines. 15:32 It's incredibly frustrating to build that system by hand. 15:35 You have to do a lot of work and be very careful, and do things like calculate the hue between the colors on the page, like what is the spread? 15:42 And make sure it's within bounds. 15:45 Agents do that, man, all day. 15:48 All day they will do it for you if you tell them to, and if you build into your system like an adversary that thinks about accessibility. 15:55 And suddenly that requirement is a thing that's built into the product that you build, and it's always verified as a piece of what it is that you do. 16:03 That's what I mean when I say you build the machine that builds the machine. 16:06 And the mind shift change is really just when you see a problem or when you see the outcomes not being good, your first thought is, "What was wrong with the machine that generated this code, and how do I tweak it? 16:21 Am I adding tests at the end? 16:23 Am I adding more adversarial? 16:24 Am I adding more of this? 16:25 Am I add… Is it, was it the architecture was wrong?" "What was it?" And then you attack the machine. 16:31 I really like that. 16:32 And I think the thing you said in there, a couple things that really jumped out was like one point of view, right? 16:36 'Cause I think this is where people get hung up. 16:37 A lot of times people feel like, I don't know, sometimes feel like they're lacking creativity or that you're just turning everything over to the agent or you're not doing anything, and it's there are, I don't know, basically an infinite point of views on how to build things, right? 16:49 And what should be built. 16:50 And I think that's the part I think people miss is like that's what you're really always-- That's what we're always bringing to the table, right? 16:56 We're always bringing… And I know you have a, you can talk about at length your history of config management. 17:01 Like there are like, I don't know if there's infinite, but there's lots of points of view 17:05 Yeah. 17:05 config management and you had a very strong point of view, others 17:09 Yeah. 17:09 Yeah 17:09 a point of view. 17:10 It's the only thing that's changed here is before you had to type every line of code, but now it's like you're injecting that point of view in, your latest thing. 17:18 So 17:19 Yeah, Yeah yes, absolutely. 17:23 So two things. 17:24 One is people are like, "Hey, I'm not an engineer anymore." And that's, absolutely not true once you start building the machine that builds the machine. 17:32 If what you're doing is just using AI inside the same structure and workflow and SDLC that you've had for years, it's awful. 17:41 It sucks, 'cause you took the fun part out of it, and now you don't do that part anymore, and now you're just stuck reviewing like the slop cannons of other people's work that comes to you. 17:51 It's no fun at all. 17:52 And but once you start working this other way, what happens is what's left is the space for conversation. 17:59 So when you ta- when you hear about people like having 18 agents running in parallel working on the same code base, like that's happening all the time at SwampClub. 18:08 We hired two new engineers. 18:11 Our average engineer, there's-- We have three who are working full-time on the product, right? 18:16 They average 35 to 45 pull requests a day. 18:20 There's no correlation to lines of code to number of pull requests, so sometimes what they're landing is a huge refactor, sometimes it's a big new feature, sometimes it's nothing but small bug fixes. 18:30 They reached that level of performance in a week they got hired. 18:36 A week later, that's what they were doing. 18:38 And the reason is, during that week, they were talking to us and to users about what Swamp was. 18:43 They were like, "What is this thing and why is it cool, and what should we do and what can happen next?" And that creativity unlock is incredible because they didn't have to be thinking about how am I gonna write this code?" And "Is Adam gonna review the architecture the way I want it to?" All that stuff was built into the system, and what was left was the space and the time and the, energy to have those conversations with each other, to spark that creativity, to push that thing out in the world. 19:08 And it's caused there's a whole game built into Swamp of how-- that's like a leaderboard that tracks everything you do. 19:16 Every time you use Swamp, you get points, and when you blog about it, you get points, and we give you like rankings. 19:21 It's super fun. 19:22 It's attached to the Discord. 19:23 People time it. 19:24 There's ways you can get like accelerators, so the whole community gets double experience points. 19:28 Like it's cra- you would never have built that shit in the before times because there's no way you could ever get it past a product person, right? 19:35 If you'd asked me as a product person, "Hey, should we build that into this product we're building?" I'd be like, "Absolutely not. 19:41 What a waste of effort and time." But Nick Steinmates had a vision of how cool that would be and how great it would be for the way the product works, and he built it and we tried it, and it was sick. 19:54 And like that only can happen because that creativity flywheel exists. 19:58 Do you know what I mean? 19:58 Like he couldn't have done it before. 20:00 There was no way. 20:00 And so like I really see this as this massive creativity unlock for everybody. 20:05 Like our ability to see what you think is good and find out if it's good has just-- it's, the time has shrunk to nothing. 20:14 What feels like nothing. 20:16 It's such a, it's such a blessing. 20:18 And getting there is real work because in between is actually this trough of awful where it's like you were great and you knew what to do and it was fun and you were creative and you were writing code. 20:31 Then in the middle, you're not writing code anymore, but nothing else has changed about your life or your day or how you work, and it's just And then on the, far side of it is actually this other place that is fantastic. 20:46 But you have to finish going, you have to keep going through that journey, and everyone I know has been on that journey to some degree or another. 20:53 Like we've all suffered that, that arc. 20:56 No, I think that's exactly right, and I feel maybe people that are, like, at the bottom of that, that ravine, right? 21:02 There's hope, and I guess that's maybe this is, that's what this episode is like, "Hey, keep going. 21:05 You can get to the other side." Because the the thing you said in there too that also I think is a common objection I'm gonna call out was you said adversarial, an adversarial agent. 21:14 'Cause one thing people always say is what about You picked usability, but that's a good one. 21:18 But people will be like I don't know how we're gonna do usability." Or sometimes the popular one I hear is security. 21:23 It's like, "I don't know how we're gonna security test." And I think that mindset of "Oh, I'm actually… I can do that." before it was like this is another group, like security. 21:33 And it's like I can sit down, open up another session, or maybe in a second we'll talk about Swamp Club, maybe we'll do, you'll tell us how to do that. 21:39 But it's and I can just tell this agent to be like, "Okay, listen, here are all the security requirements." I don't even… In fact, I can even start with "You tell me what is the world-class security requirements for this application I'm building, and then I want you to go test everything, and I want you to output a report of everything." And then maybe that agent implements or maybe go to another agent, go to the other agent and be like… And and that way they're implementing it, and you have another test 22:04 Yes 22:05 I think that's just Like I feel so strongly peop- like I don't know how to inject that into people's like, brain, 'cause they like, they 22:13 Yeah 22:14 and then they just stop. 22:16 And I'm 22:16 Yes. 22:17 No, keep going 22:18 N or command N, new terminal, ask it, ask the thing to do the thing that you're worried about, right? 22:25 Yes 22:25 don't know. 22:25 Do you have any thoughts on 22:26 Yeah, for sure. 22:28 Again, this is baby steps to how you get to software factories and how you get to changing the shape of the whole SDLC because you have a security posture for your application. 22:38 There's some things that are true broadly about security, right? 22:41 There's some things that are true about your application. 22:44 If you go ask a security team what you should be doing, like you should be doing a threat analysis. 22:48 You should be understanding what the security posture is of your specific application, where the risks are, and how you mitigate them. 22:55 All of those things are things you should encode in an adversary that then runs every single time that you write code on that application. 23:04 And so if you think about how we did it in the before times, we did it because we were slow, and we tried to rely on us being slow and talking to other people and internalizing the shape and code review and security review and all of those pieces. 23:18 That was how we tried to align our output to the security outcome that we were looking for. 23:24 Now, you take that same expertise and you encode it as part of the process, and now it just happens 100% of the time. 23:32 Is it perfect? 23:33 No. 23:35 It was never perfect. 23:36 It wasn't perfect before. 23:37 What you're looking for is a level of quality that's high enough, that is strong enough that you're like, "Yeah, this is a- at least as good as it would've been if we'd done it by hand. 23:48 It's probably better. 23:49 It's probably heaps better, actually." And if you think about how do we then improve that system if what it does is put a security vulnerability into the world or build the software wrong, what's our response? 24:04 And the answer is both patch the security vulnerability, obviously. 24:08 Also, improve the machine that built the machine. 24:11 Why did the machine allow that security vulnerability into the world? 24:15 Are we adding new black box tests to prove that we won't regress on that security vulnerability? 24:19 Are we adding new adversarial instructions so that when we're doing adversarial review of plans and code, that the adversary can check it? 24:27 Because it can test for things that you couldn't encode deterministically, right? 24:30 Accessibility is such a good example. 24:32 Usability is such a good example because it's really hard to deterministically test for those things. 24:37 There's a bunch of rules that are kinda arbitrary and that the system just has-- You have to just work out for yourself. 24:44 Same's true in security. 24:45 Like what is that posture? 24:47 And so like that's how you do it. 24:49 And to get there, back to if what you're doing is firing up Cursor or Claude or whatever, you got VS Code open and it's like writing some code for you, it's pretty hard to see how you do that all of the time. 25:02 And then when you're responsible as an engineer to do it all of the time, that's also annoying because it's like, "Ah, did you remember to run the ARP, pop another terminal and run the security reviewer 100% of the time?" next thing you're putting all that stuff together in a pipeline. 25:16 You're putting it together in behavior. 25:18 You're doing whatever harness engineering, and like software factory design, like whatever you wanna call it and that loop starts to work. 25:29 And that's that's how, you get to the other side. 25:34 That's how you get to the good part 25:36 Love it. 25:36 So let's get into sw- Swamp Club a little bit now, 'cause I wanna… I think hopefully we've convinced people they should come on this journey with us. 25:42 Now 25:43 It definitely should 25:44 the I feel like compelled as a podcast interviewer, I need to ask what's the origin of the name Swamp 25:50 Okay. 25:50 Club, 25:51 yeah, so we were so five of us who had worked together at System Initiative, when System Initiative failed part of what we saw was that we had just finished like a big POC at one of the world's largest banks, and we had won, which was all we wanted for this product we'd been building. 26:10 We'd been working on it for six years. 26:11 We finally got someone to be like,"This is sick, and it will change our lives." And it was because we had hooked it an agent up to the system we had built, and we were driving it over MCP, and the results were awesome. 26:24 But when we looked at it, we were like, "Oh, man," there was a lot that was cool about this, and none of it was the technology we'd been building for six years like, we knew how the underlying system worked, and we were just like, "Ugh, it's just not, it's not right," like it's super cool, but it was built for the wrong era. 26:41 And so I asked, we asked basically five of us who to, to… We had some capital left. 26:49 Our investors were generous enough to let us basically use that capital to see what we could do. 26:55 So we shrank the team down to just five of us, and we looked each other in the eye and said, "Okay, we're gonna go be absolute AI maximalists." this is a long-winded answer to your question, but I promise I'll get to the actual one. 27:08 and and we're just gonna see what we could build. 27:11 And so we came up with something we wanted to build. 27:13 We started to build it. 27:14 We had to give it a name, and and so we were talking to each other, tossing around names. 27:20 My requirement was that we give it a code name because you're gonna wanna throw that code name away because it's probably gonna be bad. 27:27 And we called it Swamp. 27:30 And the reason we called it Swamp was it's the, it's an anagram for the initials of the people who created it, so Steinmetz, Watson, Adam, Mihir, and Paul. 27:38 Swamp. 27:40 So then we took it to that conference I was talking about earlier, and we showed it to people, and it was called Swamp. 27:44 And we were like, "This is a prototype. 27:44 This is not a thing." And everybody was like, "That's a sick name. It should totally be the Swamp." And then Nick Steinmetz built this incredible design for how Swamp Club works and there's like swamp monsters, and you're like, when you're level five, you're like a murk lurker. 28:03 And there's all this fun stuff you can do in Swamp. 28:07 And it turns out that part of what you get back when you start building with AI is the time to engage in fun shit. 28:13 yeah, it's Swamp Club because because it turns out that Swamp is fun, and Swamp Club is also awesome, and what you wanna do is come be a part of the club. 28:24 And if it was called like Stuffy Enterprise Software Club, you'd be like, "No," "I don't wanna do that." and so yeah, we called it Swamp, and it stuck, and I love the name. 28:34 But yeah, that's why it's called Swamp. 28:36 I think you've, you took your name and I won't steal it for everyone. 28:40 Everyone should go to Swamp. 28:42 So here's the way, it's swamp-club.com. 28:44 So remember, put the dash in there. 28:46 And the website is definitely not stuffy corporate enterprise. 28:49 It's 28:49 It is not 28:50 It's like a… It has energy. 28:52 So 28:52 Yeah, it's like cyberpunk meets Miami Vice meets Swamp Thing 28:56 Yeah, there you go. 28:56 That's I, you, that's a very… I can tell you've described it before. 28:59 That's really good. 29:00 But you should check it out. 29:00 Everyone listening pull it up on your phone or on your machine. 29:04 You'll definitely enjoy it. 29:05 So okay, so let's I kinda take it there's this big idea I think of behavioral centric. 29:10 You kinda talk about an, an idea of going from if you will the agents doing everything to rebe- repeatable behavior. 29:17 So kinda take us inside the swamp, if you will. 29:20 Give 29:20 Yeah, for sure 29:21 how it all works 29:22 Yeah. 29:22 So if you've built with AI before, right? 29:24 Then you've absolutely had this experience where you ask it to do something and it does it, and you're like, "This thing will change my life." And then you ask it to do the exact same thing again, and it doesn't do it the same way at all, and you're like, "This thing is the dumbest thing I have ever encountered in my journey." Like, why? 29:39 Why? 29:39 it doing this again? 29:41 And so that's rough. 29:42 That's the s- that's the core of what Swamp solves for you. 29:45 And so the way that it does that is let's actually spend a… Since your audience is filled with nerds, let's like stay nerdy for a little bit longer. 29:55 The way that we have been thinking about building agents so far in the industry has been that you have a control loop that's con- that is the agent itself, which has a bunch of tools. 30:08 So the LLM is making decisions about what to do by having a list of tools that it understands how to use. 30:14 You give it a prompt, and then as it does more work, it's feeding back into itself and deciding "What's the next thing I should do?" So it's in control of the control flow. 30:24 And so if you look at why a lot of the workflows we build are slow or complicated or expensive, it's because that loop is sitting in the control flow. 30:36 The way it passes information is through context, so a bunch of text comes out of a program or comes out of the tool call or comes out of whatever, feeds into the next turn, right? 30:44 Making the whole thing longer. 30:46 And then we make it more complex because we have to parallelize, so we build sub-agents. 30:51 Those sub-agents also have their own context. 30:53 How do they pass information between each other? 30:55 I blat out a bunch of text at you. 30:57 And then they're trying to be controlled. 30:59 All of them are non-deterministic, so they're winding their way out. 31:02 One way you try to solve that's with skills. 31:05 Turns out skills are just context, so helpful, but also it still doesn't do what I wanted it to do. 31:11 You build a skill that has a 15-step process in it. 31:14 It follows five of 10 one time out of 100 and that's a very annoying time when it decides just not to do what you want it to do. 31:23 So what Swamp does is take that same idea and then move that control flow back into deterministic code. 31:31 And so when you say, "Hey, I want you to build me that application that sends out an email for party planning," that you described earlier, you could just tack on the end of it, "comma, with Swamp." And what it would do is break that down into a set of reusable behaviors. 31:48 And a reusable behavior is, think of it like an object. 31:51 And so it has properties. 31:54 It has methods on it that you can call. 31:56 It has typed in… Those are typed inputs and outputs, so it has schema that it understands. 32:00 It writes its own data, so whenever it needs to record information, that's also typed structured data or it's an artifact with some metadata. 32:08 And then those behaviors become this low-level encapsulation of you doing something. 32:14 For example, "Draft me an email." so you'll take into account, for example What is the email about, right? 32:22 And who is it for? 32:24 And have we sent email to the audience before? 32:27 All of those pieces would be behaviors in Swamp, and all of the data about every run that's ever happened would be stored in Swamp, and you could query it. 32:36 So you could be like, "Hey, when I get to drafting it, go look the last-- go look at the last email Brandon thought was good about a birthday party, and then use that as input, and then look at who we're inviting. Is it a children's birthday party or a grown-up's birthday party?" And instead of having those things be LLM evaluations, they can be steps in a workflow that gets stitched together by these behaviors. 32:59 And all that stuff gets written by your AI agent. 33:03 So no one's ever written a workflow by hand in Swamp, 'cause why? 33:09 Like instead you just say, "Hey, here's the thing that I want. Go build it." job is to turn that into those reusable building blocks, stitch them together, and then use them. 33:18 And so the, Magical. 33:22 So it's great at when you point it at like internal systems, custom APIs, stitching things together, like it'll just build whatever the workflow is. 33:31 It'll write the integration if it doesn't know how to do it. 33:33 It can extend itself at every layer, it becomes this building block you use to build this corpus of behavior that's what you build and scale as you build more and more complex AI systems. 33:46 So what's the relationship then? 33:47 'Cause as you were talking there, I was thinking in my mind, I was like you have APIs, then you have MCP servers that sort of encapsulate… Yeah, and depending on how you do it, they encapsulate some level of be- deterministic behavior, right? 34:00 It could, yeah 34:01 That agents use and can use. 34:04 So it's like what's the-- like, where's Swamp in that hierarchy? 34:08 If MCP is like it'll publish what it can do, and 34:12 Yeah. 34:12 has to decide, 34:13 Yeah. 34:14 in that hierarchy? 34:15 Yeah, so MCP is basically a way of adding tool calls to an agent. 34:21 So if you think about that loop I described earlier, MCP is like a tool call over a network pretty much, right? 34:29 The, and the agent is still in control of that outer loop. 34:32 So if you think about like how deterministic is the result, the answer is not very, right? 34:36 Because maybe I'll call it, maybe I won't. 34:40 Maybe I will invent some data, maybe I will not make that call and reuse something else I had from context at an earlier session. 34:47 Like all of those downsides are still there with MCP. 34:52 Where Swamp fits is underneath that layer. 34:55 So if you think about like plugging it up into like code, Claude Code or Codex or Cursor or whatever, or into a custom agent that you're building for yourself, right? 35:05 Like for example, we have one of our earliest customers built a marketing application where they allowed their marketing team to do segmentation where they pulled in data from Salesforce, they pulled in data from their proprietary targeting systems, then they pulled in data from the marketing analysis systems, and now marketers can go through and decide like how to build interactive targeting for campaigns. 35:27 And they do that in a chatbot, right? 35:30 And they built that thing on top of Swamp, right? 35:32 And and all of that stuff becomes Swamp behaviors and workflows. 35:38 So mostly it's a replacement for things like MCP because you don't really need it anymore because the agent can just call Swamp and call into those behaviors and do it. 35:48 Sometimes you wind up wrapping MCP servers with Swamp because what you want is a deterministic way to access that API and to store the results and then see them later. 35:59 And so you wind up like, that can be a very effective thing to do with Swamp. 36:03 And it's one of the things that's most fun about Swamp is that Swamp just doesn't care what crazy shenanigans you need to get up to. 36:09 Like you just, you tell whatever, you tell your agent what you wanna build and tell it to build it with Swamp, and then Swamp will figure out how to decompose it into these reusable behaviors. 36:17 Okay. 36:17 And then what makes it, like what's the secret sauce there like to make it deterministic? 36:21 So it's so before MCP, like you said, you outlined it perfectly. 36:24 It's like sometimes it happens, sometimes it doesn't. 36:26 But like in Swamp world, it is gonna be term- deterministic. 36:30 But like how? 36:31 How does that 36:31 Yeah. 36:32 So there's not a ton of magic to it. 36:35 We turn it back into code again. 36:36 Okay 36:37 So what happens is we're factoring it as code where there are typed inputs and outputs which the agent can reason about, and then we stitch those together into workflows, and think of a workflow as the control flow. 36:50 So you've got that 13-step skill you wrote that's "Always do these thir- 13 steps." And you would just take each of those steps. 36:58 Those steps would become reusable behaviors. 37:00 They have typed inputs and outputs. 37:01 They store their own data. 37:02 They can pass information between each other. 37:04 They can scale in weird ways if you want them to. 37:06 They can fan out to multiple fleets. 37:08 They can do all that shit. 37:10 And all that the AI has to reason about is, "How do I write this one behavior? 37:14 And then how do I stitch that together into a workflow?" And then I call that workflow from the AI, and everything from underneath that, it's deterministic because it's just source code again. 37:24 It's just software like you've always understood it. 37:26 Which is great because now I can refactor it, now I can measure it, I can scale it, I get all those properties back. 37:32 The, And then sometimes the step that you wanna run in that workflow is non-deterministic. 37:39 Sometimes it's like, "Hey, agent, figure out what black box test should run based on the changes that was being made in my source code. 37:49 And call Jev," which is like a system one model, "to give me a probability of whether or not any given u- accessib- any given black box test is re- would regress based on the code that was changed. 38:01 And anything that's over 80% likely to impact the code base, I want you to run." And that's a non-deterministic step. 38:09 It'll… What that thing returns will be a list of tests to run. 38:13 It's gonna choose differently every time, but you don't care because you've constrained that thing to just what you wanted it to do. 38:20 So instead of calling, letting the LLM figure it out, you're calling out to intelligence and being like,"Hey, in this moment I need smarts, and here's what the inputs are and here's what the outputs are." That's how we do things like reduce like a code review workflow, like by like a factor of 10X in terms of its token cost. 38:37 Because you stop spinning all these sub-agents, you stop communicating in all these insane ways, and you slim it down to what is it I actually need the LLM to do?" And then you just do that. 38:47 And then everywhere you can just use code, you just use code. 38:50 Which it turns out is most places. 38:52 Yeah, I was gonna say, so that's really where the token savings comes in. 38:54 So you're basically doing this handoff. 38:56 It's like, okay, when I-- when the agent is done its non-deterministic work, I will take over, I being Swamp, and then I will do this deterministically, and then… 39:05 Yeah 39:05 code, which is not gonna use tokens, which is gonna result 39:08 Yeah. 39:09 Yeah, and that behavior, like where the, like if I'm calling out to an LLM, it has structured inputs and outputs. 39:15 So I'm like, "Hey, LLM, you're gonna return me this information. Here is the schema it must match." Then I write that data out with that schema. 39:23 That's now queryable in Swamp forever. 39:26 It's stored as a versioned artifact. 39:28 We keep copies of it. 39:30 If you write more of them, we'll store the old ones. 39:32 They have lifetimes and garbage collection and all that stuff, and you can query it. 39:36 You can be like, "Hey was…" I, just had a call earlier today with someone who's using Swamp to do reproducible science because he can query every single variable of every single experiment that he ran through the lab because it's all encoded in behaviors. 39:53 And and so now he has perfectly reproducible science, and he can prove it not only like by writing it down and pasting it in a notebook literally, but by saying, "If you have the same equipment and you just run this workflow, you will get the exact reproduction of the experiment as I ran it with every variable intact." 40:11 Got it. 40:12 And then, so you've been talking about behaviors, but I also wanna there's ex- extensions. 40:17 So 40:17 Yeah 40:18 on the website here, it's I don't know, there's a lot of 40:21 Thousands 40:22 It has a little scoreboard. 40:24 So what's the relationship here? 40:26 What's an 40:27 Yeah, 40:27 What's a behavior? 40:29 Yeah, so an extension wraps up a bunch of behaviors, workflows and every other piece of Swamp that can be extended. 40:38 So all of Swamp is extensible. 40:39 So one of the things we learned in building Swamp was that you-- the future is software that is malleable to your demands. 40:46 So like for example, that data store that we write to, most people are writing it by default on your laptop, it's just writing to files on disk. 40:54 You can also then use like an S3 plugin extension, and what it will do is write that data to an S3 bucket for you. 41:02 And now the team can use that S3 bucket as a central source of information and scale it that way. 41:07 Some people also want that data in a database, and so you can write in a Swamp extension that moves the data store from an S3 bucket into Postgres and store it as blobs and TOasts. 41:18 Or you could store it in MongoDB or wherever you wanted to put it. 41:21 So it's extensible sort of all around its, the behavior, because in the world of AI, you can just, if it does, if a product doesn't work the way you want it to, you'll just rebuild it. 41:32 Mhm. 41:33 And so I want you to keep using Swamp, and so I have to make it so Swamp adapts to you, not the other way around, right? 41:38 It can't be you adapt to Swamp, it has to be Swamp adapts to Brandon. 41:42 And so extensions are the way that Swamp adapts to Brandon, and it's the way that you then share that behavior, those things with other people. 41:49 So like a good example there is like we have full coverage of like the AWS and the Google and Azure APIs, and so if you wanna do work on those platforms, there's just extensions there that wrap'em up, that have behaviors that just do it, and you don't have to think about it. 42:04 It, the, they already work and we maintain them for you. 42:08 But maybe you want it to work a little differently. 42:11 Maybe Brandon's version of EC2 has a bunch of other things that you would like to wrap around it, so you can write an extension that literally extends my extension. 42:19 You can be like, "Aha, but Brandon EC2 has this new magic method attached to the EC2 instance that does something specific that you want it to do." and that's the magic of extensions. 42:32 Yeah, and I was looking here, so it's like you, you being swamped, you've written obviously a bunch of these, and then looks like some people, other people have done them, and so they 42:39 Oh, yeah. 42:39 The community writes a ton of them. 42:41 Yeah. 42:42 And you get a bunch of points when you do it. 42:43 So when you do it, it rockets you up the leaderboard writing 42:46 Yeah 42:47 and publishing them. 42:47 Yeah. 42:48 When we announced our fundraise, one of my favorite things that happened that day was we had a community member who had built a bunch of cool extensions, and he had batched them all up so that he would-- 'Cause he, what he wanted to do was get to the top of the leaderboard. 43:00 So he waited to publish them 43:02 Right 43:04 until, He didn't know that it was the day we were announcing our fundraising, but he started a campaign where he was posting little poster images into Discord being like, "At this time, I'm gonna trigger the double points accelerator." And so he told everybody when he was gonna do it. 43:18 He triggered the accelerator, then he published 15 or 20 extensions, and it rocketed him straight to the top of the leaderboard because he got double points for every extension he published. 43:29 Nice. 43:29 fun. 43:30 There's a lot. 43:31 You can see there's a ton of stuff that has huge numbers here. 43:33 I just look at the s- the scoreboards and stuff like… All right, so as we wrap up here, let's-- I always wanna leave people with actionable stuff. 43:39 It's like, okay I know on, and I know one answer is it's easy to install. 43:43 You copy this little curl command, and 43:45 Yeah, it's pretty… Yeah 43:47 that, what's… Okay, so I'm ready to get going. 43:49 I wanna embrace this lifestyle. 43:51 I wanna if you will, the machine 43:53 You're gonna join-- Yeah, you're gonna join the swamp club 43:55 to. 43:56 Okay, so I got Swamp installed get me going. 43:59 Like, how do I 44:00 Yeah. 44:01 something happening? 44:02 Yeah, you're gonna run Swamp start and then it'll guide you through it for-- It'll just guide you through it. 44:08 It'll have you-- Basically, it'll pop up a web browser. 44:11 It'll ask you what you wanna see or what you wanna build. 44:14 It'll ask you what harness you wanna use, then it'll pop a terminal window with that harness in it with that prompt, and it'll have already set up Swamp for you, and you can just see it start to work. 44:22 And the best way to start is to just think about something you want and just ask it to build it with Swamp. 44:29 Literally, if all you did was pop up Cloud Code and be like, "I want to build a party planner," blah, blah, blah, blah, comma, "with Swamp," and then hit enter, what will happen is it will go install Swamp, and it will initialize a Swamp repo for you, and then it will figure out how to go build that thing. 44:45 And, from there, you'll iterate. 44:47 You'll just, you'll be back in the loop of "Do I like what's happening? 44:51 Do I understand how it's fitting together?" "How can I grow and change the way that this thing works?" And that's how you get started. 44:59 The other thing that I would really recommend people do is come join the Swamp Discord because the number one thing you have to do in this moment is just talk to each other. 45:10 The more that we talk to each other about what we're building, about what works and why, what doesn't work. 45:16 Software factories, for example, still really weird, the shape of what we're coming up with is different and unique to everybody who's building one, so we have to talk to each other, so we start to learn what are the patterns that work? 45:29 And why do they work? 45:30 And how could we improve? 45:31 And no one is actually behind in this moment. 45:34 Everybody thinks they're behind, but you're not. 45:36 Whatever you're learning, there are thousands of other people at exactly the same moment that you're at, and you just need to talk to those people. 45:44 And then there's other people who are a little ahead of you or who have already experienced it, and you need to talk to those people, too. 45:49 And the Swamp community is great for that. 45:52 Just come and hang out and be like wherever you're at in that journey it's all good. 45:58 And there'll be this huge crew of people who are just figuring it out together, and they're having the best time. 46:03 All right. 46:04 I think that is a v- powerful call to action. 46:06 I would even add to it, it's challenge yourself to just bring your point of view. 46:09 That's what we talked about. 46:10 It's like you start up Swamp, and that's what you're gonna bring. 46:12 I wanna build this thing, and this is my point of view, comma, I guess it was comma Swamp, is 46:16 Yeah, comma with swamp 46:18 With Swamp, comma with Swamp, and then just see, challenge yourself. 46:22 How far can you go? 46:23 Yeah 46:24 I think this is what I always tell people and like, I always tell, like I feel like AI, when you get to it, it's a, like you feel like you have these superpowers. 46:30 But until you experience the superpowers, like you're super skeptical. 46:34 And 46:35 Jetzt? 46:36 Ja 46:36 trying to tell people like just just tell it to do the thing you really wanna do. 46:40 No, don't write a script, have it do the thing, the big 46:43 Yeah. 46:44 and see if it happens for 46:45 It's just very difficult to get to the part where as a professional software developer or professional operations person, to realize that now what you're doing is talking about outcomes instead of talking about implementation details. 46:58 It's a trip. 46:59 And but once you start to experience how to build those systems and take the baby steps you don't have to jump all the way to the end. 47:07 The end is a lot. 47:10 And and there's a lot of cultural stuff. 47:13 There's a lot of inter- how you interact with people changes. 47:16 There's a lot of change that happens here, and it's okay to just take that change in chunks, it's okay to take a little bit and then move forward, and take a little bit more and move forward, and take a little bit more and move forward. 47:26 You don't have to do it all in one day. 47:29 And it feels like you do because it's moving so fast and the industry is constantly telling you that if you're not on this train, we're all gonna die. 47:34 But it's gonna be fine. 47:37 We're gonna figure it out. 47:38 It's gonna work out. 47:39 Just take what you can do and then go touch grass. 47:43 And and then tomorrow come back and learn a little more. 47:46 And we The Swamp Discord is a great place for that. 47:52 Your podcast, this episode's a good example of it we're gonna learn from each other what we should build and how to build it and what it means to do stuff in this era. 48:01 And it's not a moment to wait for other people to tell you what to do. 48:06 It's a moment to figure it out and to explore and to build because that opportunity is here. 48:13 And it doesn't-- It's not always, we spent the last, especially like in infrastructure, we spent the last 12 years, 15 years basically in stasis. 48:22 Nothing interesting or new has happened. 48:25 Nothing. 48:26 We're doing it exactly the way we were doing it before. 48:28 And not true now. 48:31 It's a fun time. 48:32 Yeah, absolutely. 48:33 So listen, okay, for everyone, I'm gonna put links to the Swamp Club, the website, the Discord. 48:37 I'll put Adam's LinkedIn stuff to some of his other websites. 48:41 And then Adam, I believe, are you also in the podcasting world now as well? 48:44 Did you, So 48:45 have a podcast. 48:47 Yep, so we do, we have a podcast. 48:48 You can go to swamp-club.com/podcast, and what we do is read you the news. 48:53 So if you're burned out, like reading Hacker News and having to keep up with all the AI news and all the we talk a little infrastructure, we talk engineering, we talk venture capital, and we talk AI. 49:05 And it's basically me, my two co-founders, Paul Stack and Nick Steinmetz, and then a special guest. 49:11 You should come be a special guest, Brandon. 49:12 That'd be super fun. 49:14 And we just talk about the news of the week like we're Muppets. 49:19 And so we break it down for you, and then we give you hot takes. 49:22 And so you can just spend an hour every Monday eating a snack and hearing us tell you all the news so that you don't have to go read all that shit all day 49:32 Amen. 49:33 Of course we're huge p- fans of podcasts as well as producing podcasts. 49:37 All right, Adam, I've learned a lot. 49:38 Been an excellent episode as always. 49:39 Thanks for coming on the show 49:41 Oh, it's my pleasure. 49:41 Thank you so much 49:43 All right, and thanks for everyone listening. 49:44 And if you feel so inclined always give us a review five stars. 49:48 We'll, never turn that down. 49:49 And with that, we will talk to you next time Thanks for listening. 49:54 Check us out at theenterpriseaishow.com for past shows, newsletters, and all things enterprise AI
Community discussion