# EP. 01 · Your job was never to write code

- Show: After the Last Engineer (ATL/E)
- Guest: Vladimir Petrov, Cobalt Code (https://cobaltcode.ai)
- Released: 2026-10-01
- Length: 54:40
- Watch: https://www.youtube.com/watch?v=8ko36BavICI
- Page: https://atle-site.netlify.app/episodes/01-vladimir-petrov/

Vladimir pointed agents at three months of open tickets. They opened 25 pull requests at once and took a server down. Then the bottleneck kept moving, from review to CI to the one person approving each deploy.

About the guest: Has written code for 30 years. Builds a software factory at Cobalt Code.

## Where they keep the lights on

Rooms lit: What to build. Darkness level L4.

> You need to end up with someone that owns the product to say, what we want and what we don’t want. [39:10]

## What broke

- [25:46] Agents opened 25 pull requests at once and took a server down.
- [5:45] An agent left alone for two days. Code that didn’t work, tickets that didn’t matter.
- [2:16] The bottleneck moved: review, then CI, then the manual confirmation on every deploy.

## Quotes

Each quote matches the transcript at its timestamp.

- [3:23] "They are not hype."
- [28:07] "We like to call it agile. It’s quite waterfall."
- [5:45] "You need structure around agents."
- [22:24] "You can keep the stuff you love, and you can just give away everything else."
- [53:37] "Everybody’s experimenting with software factories, but I don’t think anyone has the right setup now."

## Transcript

Machine transcript from Descript. Speaker labels and words can be wrong. Quote from the Quotes list above, or check the timestamp in the video.

[0:00] **Vladimir Petrov** If you don't give them proper instruction, they go and just do random, you know, stupid crap. Everybody. I scared them a lot because I told them that all of us will probably have to change the way we do our job. People were resisting to throw away code. It takes five minutes to build another version.

[0:16] **Vitaliy Filipov** Just tokens.

[0:17] **Vladimir Petrov** Tokens are expensive.

[0:17] **Vitaliy Filipov** Have like a forecast when this will happen.

[0:19] **Vladimir Petrov** God, I don't know. I think that if you say five years, this is too far ahead in time. What is the factory of the future going to look like? It will be fully automated and there will be one person there with a dog, and his only job would be to feed the dog.

[0:36] **Danail Arapkuliev** More software used to mean more engineers. That was the deal for 30 years. Now everything is different. Agents write the code, and nobody needs to read every line of it. Vitaliy and I started this show to understand what, what changes when agents do all the work, what changes for teams, for companies, for us.

[0:59] Building used to be expensive, then we learned how to validate first. AI made building cheaper. A factory make it cheap enough to throw it easily away.

[1:10] **Vladimir Petrov** Perhaps we can simply create a version of the product. If it's something they don't want, you throw it away, 'cause it's all code.

[1:16] **Danail Arapkuliev** That validation on steroids.

[1:19] My guest is doing exactly that. He's writing code for more than 30 years. Enterprise work where nothing move fast. He started by cloning Lovable just for fun, but big clouds were too heavy to run it, so he started building his own sandboxes.

[1:35] **Vladimir Petrov** You can't have a container up and running in seconds and then tear it down and then create another one, or create 100.

[1:40] **Danail Arapkuliev** At first, his team didn't want it.

[1:43] **Vladimir Petrov** Don't let my LLM touch my code base, my precious code

[1:47] **Danail Arapkuliev** base. So one day, without telling anyone, he just

[1:53] **Vladimir Petrov** plugged it in. It just blew up and just created 25 PRs all of a sudden. It actually bought one of the server.

[1:57] **Danail Arapkuliev** So that became Cobalt. The idea is very simple. You don't usually share your PC with a teammate, so why should you share it with an agent then?

[2:07] **Vladimir Petrov** QAs don't need to go to developers anymore. They have developer. It's the agent in the code.

[2:13] **Danail Arapkuliev** But the bottleneck didn't disappear. It moved.

[2:16] **Vladimir Petrov** I was the next blocker, and then the biggest blocker was actually the manual confirmation when it deployed.

[2:22] **Danail Arapkuliev** In this conversation, we will talk about why agents fail the same way people do.

[2:27] Should we build or buy software? What happens when a team stops being afraid to throw away code? And where humans stay when the factory runs. This is Vladimir Petrov, founder of Cobalt Code. I'm Danny, and I'm here with Vitaliy, and this is After the Last Engineer. And just one more thing. This is our first episode.

[2:50] Vitaliy and I are trying to figure out what software dark factories means for people like us. If you are asking the same questions, just subscribe. Let us figure it out together

[3:01] I entered a kind of channel where my colleagues had their agents there, and I entered with my agent and me as a human and started doing some wild stuff together with the different agents. So my first question to you, Vlad, is that real or hype, dark factories and software factories? What do you think?

[3:23] **Vladimir Petrov** Well, the software factories are here to stay for sure. They are not hype. The problem is that, well, there is obviously a lot of hype around them right now. Everybody is doing it themselves in their way. Everybody is trying to figure out how, what's the best way to do it. I just feel like that, if we calm down a bit and came back to basics, because agents are so much like people.

[3:48] I found out that if I treat them like coworkers instead of this computer thing that talks to me, it's easier, and when you manage them like people, it's actually easier. Obviously, they work twenty-four/seven, and you can spawn as many as you want and all of that, you know. So you end up controlling armies, you know, at some point.

[4:08] but I feel like they have the same, the same deficiencies. For example, communication. If they don't communicate properly, if you don't give them proper instructions, if they don't have access to the right data, they go and just do random, you know, stupid crap. So-

[4:25] **Danail Arapkuliev** do you think they can cooperate in a way that humans cooperate, in order to produce value? Or, like they-- when they try to do something together, they're starting to... One is doing better the other, and so they're kind of comp-competitors on what, on the task they're, they're doing.

[4:46] **Vladimir Petrov** Well, it's the same if you have a bunch of people that are equal in terms of responsibilities. Because nobody takes ownership, you end up, doing-- wasting a lot of time and people just talking about what to do. This is the so-called design by committee, where you have everybody has their own opinion, everybody's opinion is valid, and then, you know.

[5:09] Well, yeah, if you end up in this situation, then the agents, they're not going to be very helpful. I actually tried it with, OpenClaw, because OpenClaw lets you create sub-agents and all of that, and I just gave it an idea I had for a piece of software, and I told it, "Yo, just go ahead and build it and let me know when you're done."

[5:28] and then it went and just spun its wheels for like two days and created a bunch of code that didn't work and a bunch of tickets that didn't matter, and, you know, a-and the chatter was, like, irrelevant. It's like rubbish. You know, they were discussing things that didn't matter. They didn't push anything forward.

[5:45] So that was, one of the failed experiments of mine. So yeah, you need structure around agents. and it f- The way I see it, you know, you, if you structure them like people, it is going to be more successful than if you try to invent new ways of structure. So dedicated teams, certain agents needs to have ownership of the task.

[6:09] They can delegate down, and obviously command and control and all of this, all of these patterns that we already know. the benefit is obviously, with people, you put someone in a position somewhere in the organization, and, they're not-- it's very rigid. You can't just move them around. With agents, it's not a problem.

[6:29] You know, you can experiment with organizations. You can experiment how you want them organized, how you want them set up, and you just delete the agent that is not doing the, you know, the work as you want it, spawn some more. That's it. So it's much more flexible in terms of, you know, getting to where you want.

[6:46] So I don't know how vague this is. This may sound vague. but, for example, when, when I started experimenting, at one of, one of my projects, I tried to give them, strict roles. So for example, you have a QA agent, you have a, you know, a coder agent, you have a project manager. And, it works to a point.

[7:09] but, again, you waste a lot of time in coordinating. you can tell that, you can see that how they, you know, coordinate a lot. so I don't know what the answer is, to be honest. I feel like we have yet to discover the most efficient way to do it

[7:24] **Danail Arapkuliev** back to your story actually. Like, from what I know about you started, a little bit as a, as a builder, for many years. You had run a consultancy for years and, also started, experimenting with AI. So go a little bit

[7:41] in time and share, what interesting experiments did you, did you hit around that?

[7:48] **Vladimir Petrov** So, I'm hands-on coder. So I run this consultancy as well. We have some clients in England and in the States, but I mostly enjoy writing code, and I've been doing this for 30 years in many organizations. and as you probably know, both of you know that it's super different when you're writing code for yourself.

[8:07] It's super different when you're writing it with your friends, and then when you do it in a startup and when you do it in an enterprise. In the enterprise, it's a completely different animal. because if you're writing code for yourself, your efficiency matters or, you know, to a point. Because you're, you're the client, so whatever you do, you know, you- you're happy with it.

[8:33] If you do it with your friends, there is still the-- you know, you can do stuff, you can take shortcuts, you can break stuff. It's not a big deal. with enterprises, the individual contribution or the effect of the individual contribution is not so much so important in a way. they look at this big structure of teams, and they want to, basically de-risk.

[8:56] So the technologies are not always cutting-edge. The processes are much, much slower than you're-- you would think. Obviously, there is lots of requirements for, auditing for all sorts of, you know, all sorts of red tape. so what interests me is, So why am I saying this? Because I think that the, agent era right now is at its infancy.

[9:18] It's-- We are getting excited by things we can do on our computers, and we are just so getting to the point where we are trying to build something that works for larger teams. and to build the right structure for this for enterprises is still out there, so I think this is a great opportunity. But when, when I started with, dealing with agents, it's when I heard about, you know, everybody was talking about ChatGPT several years ago.

[9:44] Oh, you know, it writes jokes. It can, you know, write an essay or whatever, you know, all those sorts of things. My-- I was, sitting at a restaurant and the owner, who's a friend of mine, he's this Greek guy- he's really opinionated. He thinks that, the Greeks invented the pizza. So, you know, you need to listen to

[10:04] **Vitaliy Filipov** Okay. That, that sounds

[10:05] **Vladimir Petrov** him.

[10:05] **Vitaliy Filipov** opini-opinionated

[10:07] **Vladimir Petrov** an opinion. Yeah. So, you know, if you try to tell him anything about moussaka or sarmí or any of that, he gets really, you know, this is Greek. But anyway, he was telling me about ChatGPT. Oh, did you hear that? Did you hear this? and I'm like, yeah, I know, but it's this gimmick, and yes, maybe it will become useful at some point, and I decided to look at it from a coder perspective, what it can do, how can it write code?

[10:30] And, as I was playing around with it, as you probably know, I mean, all of us have probably gone through this, we discovered that, because it's very good with language and why is it not-- I mean, programming languages should also be good with it. They're kind of, you know, they are, pa-part of what the LLM is.

[10:49] so I saw that it writes code, and I tried to make it write better code because at the beginning it was very poor at writing, at writing code. So then we started doing, you know, the-- if you beli- if you remember two years ago, there was, or earlier last year, let's say, there was a lot of, ways to how to spec things.

[11:12] You need to write a plan, you need to write this, you need to write that. There was lots of ways to make them work, and it was pretty obvious to me that these are crutches. These are crutches to things that the, as the LLMs become better, they just don't need them anymore. So I don't know about your experience.

[11:29] I don't write plans almost any-anymore, or at least I don't write them to the, to the, detail that I used to because before I would tell it everything. Architecture, I would tell it, you know, you need to code this, you need to make that. You know, look through the code, make sure it followed everything.

[11:47] Now we are talking about planning that is at feature level. So I want this feature, I want it to work like this. Here is your limits. so things are getting better and back to the, back to last year, I watched this video about the end of programming.

[12:02] So what happens when, as Vitaliy said in our previous conversation, what happens? Are we there that we can just turn off the light and leave the agents to work on their own? and unfortunately for a lot of my colleagues, I think we are very, very close to this, because...

[12:18] **Vitaliy Filipov** 

[12:19] close in terms of like, capabilities or close in terms of time? Because, if we are like 99% close in capabilities, but unlocking the last 1% is like exponentially more difficult than the current 99%, maybe not that close. But maybe if you refer to close in time, do you have like a forecast what-- when this will happen?

[12:41] **Vladimir Petrov** Oh God, I don't know. I think that if you say five years, this is too far ahead in time.

[12:46] **Vitaliy Filipov** Ah, okay.

[12:48] **Vladimir Petrov** I think it's going to be closer now

[12:49] **Danail Arapkuliev** maybe you can even start, introducing how did you build Cobalt, because people maybe don't, don't know what is it?

[12:56] **Vladimir Petrov** Yeah. I was trying to build a copy of, Lovable last year because it felt very interesting from a technical point of view to me. And I also like how, you know Lovable, the, pro-the prototype, platform that you can basically

[13:12] **Vitaliy Filipov** we use it also. Yeah

[13:13] **Vladimir Petrov** Yeah. I mean, there are plenty of clones and plenty of other-- There is open source versions.

[13:19] but, it needs several components in order for it to work properly. and one of the components is a proper sandboxing en-environment or like a orchestrator. When I started, I decided to build my own, partly because I wanted to figure it out on my own, but partly because I didn't want to be, you know, relying on existing ones.

[13:41] And at the time there was Daytona and E2B. I think these were the biggest ones. If you try to, to create a containerized application, for example, in some of the bigger cli- bigger clouds like Azure or AWS, it's not as simple. you can't thr- have a container up and running in, in seconds and then tear it down and then create another one or create 100.

[14:02] You can't do that. They have limits. it is, it is really clunky how to do it. So I ended up building it myself. So I started from Lovable clone, and then I thought, okay, maybe this can be useful in, with our teams, because our teams, we have a lot of, business analysts in our teams as well as developers.

[14:22] A lot of, several, business analysts. And what they do is they will capture requirements from the business, and they will go and have a bunch of discussions, with the business. they will write down the user stories. Then this comes down to a development team. They implement it. You know, in a couple of weeks, you end up seeing something and the business says, "Oh, that's not what we want."

[14:43] So so how do you deal with this? So we need a way to prototype quickly, and I felt that prototyping quickly, like you can do it in Lovable, only we weren't allowed to use Lovable in the enterprise 'cause it's this vibe coding platform nobody likes. So if we could prototype quickly using KI, we could potentially, you know, shorten the feedback cycle.

[15:04] So you can possibly, or this is still, not there yet. You could possibly do it as we, as you talk to the, to the business. so they say, "I want, a document library in your, in your product," and then you kind of write the prompt, and then it builds a document library. And then you say, "This is something you want."

[15:22] "Yeah, but I want the, you know, the widgets to be on the left," and blah-di-blah, "and I want this filtered this way." And everybody has their own business requirements, and when they see it running, you basically shorten the feedback cycle, and you don't have any more surprises when you deliver the actual software.

[15:38] So and this was my thinking last year. So I wanted to build something that can prototype stuff. But then you have to prototype, you have to run it on something. So I have built a sandboxing, platform. So then, you know, so then we started using Cursor in our team, so prototyping became less of a priority because our BAs could just run Cursor and create HTML pages with, prototypes.

[16:03] So we were getting there- and the next thing I built basically, by the way, the prototypes weren't very, they weren't very high fidelity. So they would be your standard HTML using, say, Bootstrap, for example. Yeah. Which means it's not using the product icons, it's not using the product, you know, like, styles and everything.

[16:25] The product that you are building a prototype for. and, so I built a knowledge agent, something that can pull this information and then can let you create a prototype that looks and feels like the product you are building it for. So we used that for a couple of months. and this was fairly successful, and it turns out that business is pretty hap- happy to have this.

[16:45] They're pretty happy to have an HTML page that they can download, run locally. They don't need servers, they don't need sandboxes, they don't need, you know, hosted stuff. This was plenty. but then, y- you know, for me, it was always a waste. You build this prototype, and then the developers have to decode it.

[17:05] Somebody has to write user stories. You know, k- testers are going, how do I test this? It doesn't look like the prototype." And then you go into the discussion, "But this is a prototype, you know, it's not the real thing. You need to make sure it makes sense." All of this. Again, lots of back and forth. So I think we are at the stage where we are actually, automating that as well.

[17:26] So people build, the BAs build the prototype, the prototype gets converted into user stories. this is combined with knowledge from the product that we are building for the client, so the user stories are relevant. And then you can feed that into a swarm of legend-- of legends, I hope, but agents.

[17:45] **Vitaliy Filipov** Agents that will be legendary

[17:47] **Vladimir Petrov** yeah. Absolutely. You can fee- feed that into, into a swarm of agents, and they just go and implement it, or implement at least the stupid, the silly, the grunt work, the stuff that nobody wants to play around with, or the stuff that is-- it takes a lot of time to be pixel perfect, to make sure that, you know, your, UI looks well in all resolutions on mobile, on, on iPad, on this browser, on that browser.

[18:12] This takes a lot of time, and agents can, can automate that So we are not there yet, we are almost there. I mean, we do a lot of development now with agents, and we use, Cobalt ex-exclusively, with our teams and with some other people. and what I noticed that, you know, the whole process changes organically.

[18:33] So before the QAs would ask us questions, so they will see a problem, and then they'll wait until a meeting, and then they'll go, "Oh, is this supposed to work this way or is it supposed to work that way?" Or is it, you know. Now they can just ask the AI for this. and the second thing is, if they need something small fixed, they don't have to wait for us.

[18:54] They can simply create a ticket and then just mark it as auto. This is our way of, letting Cobalt know that it needs to pick it up. And then Cobalt spins an agent, looks it up, judges if it can implement it, and then it just implements it. So this is the first part of the, of the software factory where it's, where it's on demand.

[19:14] People are still creating tickets, and people are still reviewing the work at the end. But in the middle, in the middle, you have a bunch of agents that just do the work. And because code is cheap, agents are cheap, if they do something wrong, you just throw it away. You s- you s- spin up another one. And actually this--

[19:32] I was going to say that this was a big, mind shift in the team. People are-- were resisting to throw away code, which is really s- really silly because I'm like, "Why? it takes five minutes to build another version. Don't, don't worry about it."

[19:44] **Vitaliy Filipov** Just, just tokens.

[19:45] **Vladimir Petrov** Yeah. And they were like, "Yeah, but tokens are expensive."

[19:48] Yeah, but don't worry about it. Just make sure, you know, we are here to deliver. it's funny how, developers and all people in our industry, are missing kind of the forest for the trees, because it's not the code that matters. it's the end product that matters. You're solving a problem for a customer.

[20:04] This is what matters. The code is a way to solve it, which means that in a few years, maybe code will be predominantly something agents write. So our job as system engineers and software engineers is not to, you know, w- write code, even though it's super enjoyable and a lot of us love it. It's actually to solve problems.

[20:26] It's all about problem-solving. And having agents and having AI and having all of those technologies in the middle, actually, it elevates us. So we are free to think about the actual problem, what we want it solved. Not the stupid bug that, you know, something is off ten pixels and it's, you know, it's annoying to reproduce or, some memory leak or an exception happening somewhere.

[20:50] This is, this is something you can delegate immediately, without thinking and don't worry about it. It will be fine

[20:56] **Vitaliy Filipov** Yeah. Th-this reminds me of, if I have to plot like our thinking model in terms of how it introduced in our teams. First it was that we just use it as a autocomplete type of thingy, like a very fast way to maybe scaffold a user test, a unit test, I mean. Then that, immediately started, okay, can we plug it to review some of the PRs and help us to spot some inconsistencies there? And at some point in time, when you pile up those type of, automated activities, you just realize, okay, we are maybe reaching a threshold, we are passing that, where actually all of the agents need to do all of the work.

[21:42] **Vladimir Petrov** 

[21:42] **Vitaliy Filipov** how was it with your team? Like, did you have this type of internal discussions like do we need to, let's say this as a software factory? was it like proac- proactive way of discussion or it was like, just following along and at some point in time you ended up with Cobalt and everything

[22:04] **Vladimir Petrov** So I wanted to push this as soon as I could, but, I got a lot of resistance from, lots of people, including people that I thought were, cutting-edge developers. People, software developers don't like to, you know, give away their, th-their stuff. They like their thing. and I get it. I love it too.

[22:24] But, you know, and the thing that agents give you and what I've tried to basically tell them is that you can keep the stuff you love, and you can just give away everything else. and for me, it turns out I, you know, I've been writing code for many years, and I still like to do it, but I much more prefer to see finished products working, running, solving problems and all that.

[22:49] so back to your question, we have, you know, in teams normally you have people that work, that, live and work in different, different speeds. and you can see that, what they noticed, what I hoped they noticed was that first of all, LLMs are not scary. second of all, they can write code that is as good as yours, as ours.

[23:11] and thirdly, it's super cheap, so you can always, you can always throw it away. You don't have to worry about this thing ending up in production and killing the world, whatever, you know? 'Cause everybody's like, "Oh, you know, I won't let, my LLM touch my code base, my precious code base," or whatever.

[23:29] There's no need to be afraid of them.

[23:31] **Vitaliy Filipov** Yeah. Did you have a pushback from the technical side or more from stakeholders that are

[23:37] **Vladimir Petrov** Everybody.

[23:37] **Vitaliy Filipov** as well?

[23:38] **Vladimir Petrov** Everybody was pushing back, yeah. I've been talking about

[23:39] **Vitaliy Filipov** Crazy

[23:40] **Vladimir Petrov** AI and how we should be using AI in our teams and in our products and everything ever since, this thing started writing code. So like for the last two years. we normally because most- all of us are remote, so we would get together once or twice a year and have presentations and like talk about stuff and where things are going and, you know, about this, and my regular presentation was about AI.

[24:07] the first one was what it, what is, what is this thing, what it can do. the second one was how we can use it in, you know, in the enterprise for coding. the third one, I scared them a lot because I told them that all of us will probably have to change the way we do our jobs. Everybody was super scared and but, you know, I always look to the, to the future, and I think it's almost there.

[24:30] and I think a lot of people that, you know, are working at the edge- Mm ... of things, although we are not really at the edge. We are the, at the edge of the things that are publicly available. I mean, there is lots of la- lots of labs that work on things that will become publicly available in five years and will blow our minds, I'm sure.

[24:49] but working at the edge of what is publicly available gives you a glimpse of the future, and, I seem to think that the future is right there, and it will come, like, next week. but it takes a while. There is inertia. There is inertia in organizations. There is inertia in people's thinking.

[25:06] People don't like to let go so easily of their, you know, ideas and the way they are used to working. So there is a ... Th- there was a lot of pushback, but I can see, actually, the biggest change happened when we introduced, when I introduced Cobalt into the team. Because previously it was all nice to have and, like, fictional, all the things that I was saying, and Cobalt came, and I just plugged it in, one day, and I told it, "Just look at the last, the, all the open tickets from the last, three months and just try to fix them," and it just blew up and just created 25, PRs all of a sudden.

[25:46] So nobody expected it. I didn't expect it either. You know, to be honest, it actually brought one of the servers down because I hadn't planned this correctly. but, 25 PRs, like this. we threw away some of them, but then we just reviewed and merged the others and, you know, y- you could tell that this is something that actually is useful because those were bugs that nobody wanted to touch.

[26:11] those were things that were like, it takes too much time, or it's good, nice to have, or whatever, you know? And this thing just went and just bulldozed through them all. Yeah. and, you know, and then I gave them a- 'cause I didn't tell them that I was going to do that. I kind of surprised my team.

[26:29] And then we had a, basically a follow-up conversation. I told them, "Look, guys, y- you don't need to be afraid of it. You can just delegate work. What you think is useful, you can delegate. if it is smaller bugs, the QAs don't need to go to developers anymore. They have developer. It's the agent in the cloud.

[26:46] Just delegate it and see what happens. Let, let's test it out." And, a few weeks later, we are, we are ke- we are doing this constantly. So people are not, don't even bother to ask anymore. You know, even if, if sometimes, for example, there is a glitch or the PR doesn't arrive, in time, they will ping me and say, "Yo, what's going on?"

[27:04] I'm like, "Go to Cobalt, ask it." So you type what happened to this PR, it tells you, 'cause it knows everything about everything. So-

[27:12] **Vitaliy Filipov** a cool, cool big bang event. we also, when we roll out, a software factory for s- some of our partners, the first time that they touch it, they usually start, "Okay, this is amazing." But I think after one month, the hype at le- at le- at least a little bit settles, and then people start to realize, okay, this is something new.

[27:35] Like this is a new universe that we are participating in. and the struggle is still there, but it's, it has moved and it's somewhere else. And for your teams, how do they struggle right now if it's not

[27:47] **Vladimir Petrov** So our,

[27:48] **Vitaliy Filipov** that?

[27:48] **Vladimir Petrov** you know, our teams are, they're fairly classical, classically structured. So you will have some developers, you'll have QAs. Most of them are, manual. some of them write obviously automation and try to automate stuff, but they're mostly manual. and, you know, so BAs will pick up, will pick up business requirements.

[28:07] They will turn that into user stories, epics, and documents, and prototypes, and all that. This gets planned into a release, so we like to call it agile. It's quite waterfall so I won't call it completely. But we try to deliver quickly and, as quickly as possible, so we have basically releases to prod every month or, or so.

[28:29] I try to push for releases every week, but, the business isn't ready because, as I said, it's an enterprise. The enterprises, they like to have... They have to de-risk, so they like to dedicate resource for acceptance testing for this, for that, sign-offs, this, you know. It-it's always a process. So, the team, how they use it, right now they will just,

[28:55] Ini-initially when I started it would, pick up everything. So every new story that comes up, it will try to build it, and this creates the, some, a little bit of chaos. First of all, there was too, too many things to review. Even though s- a lot of them were reviewed automatically, you still need to go and have a look what this thing has done.

[29:15] CI was the next blocker because all of the PRs are trying to get crammed in and deployed, so this became a blocker. You would have to wait for a couple of hours before, you know, it will trickle down to one of the environments. And then the biggest blocker was actually the manual, confirmation when it deployed.

[29:34] so our QAs were complaining the most, as you can imagine and I'm trying to figure out how to solve this, and I think that we are at the stage where, we'll need to put some visual confirmation, from agents. Because right now our agents, they do only, you know, pull requests, code, base analysis, and like reviews and stuff.

[29:55] but, we have to do it visually. So Cobalt has the ability to record videos and take screenshots, which I think is going to be the next step. We have to ask the agents to, replicate the pro- the bug, take a video, you know, narrate it, say what they are finding out about it, then, create a document of how they're going to fix it for the more complicated ones, and then present this at the end.

[30:22] So you have visual proof that things are, you know, things are done properly. and again, see, I'm thinking how to automate. I'm not thinking how to create more work for the, for the QA or for the others. this automation gives them the ability to choose what they're going to verify. sometimes you will get, situations where you need certain input data, certain account type, you know, all sorts of stuff that, you know, in order to replicate a problem and confirm it's fixed.

[30:52] and these are the gnarly, like complicated things that maybe the manual QAs would, would enjoy, doing. Or the other thing is, the agents are very good at creating, test cases. but sometimes they miss things, and sometimes they overthink. So, you know, when you-- when it comes to validation, verification, and all of that, you still need a person to, have the, you know, to, kinda, set the guidelines.

[31:18] Set the guidelines, the limits, how we wanted this tested and so on. So it feels to me that the more we i-i-implement agents and the more we include agents into our processes, we are taking a step back every time. So initially autocomplete, you know, then it was local work on your machine. now it's swarms of agents, and you're constantly stepping back a little bit.

[31:44] So i-it won't be far when we step back completely and go and feed the dog.

[31:52] **Danail Arapkuliev** share a little bit more, like, business cases from business perspective, how exactly people are using factories? Like, I mean, not from technical point of view. I can share, for example, my story, with that. So I'm v- big fan of validation of ideas. You mentioned Lovable, you mentioned, solving problems.

[32:12] So for me, like, it,

[32:14] was never, the code that solved the problem, it was the business solution that you propose. And, before, even before Lovable, my thinking was, let's create something, the stupidest and easiest to do thing in order to measure, the commitment of the people, like in skin in the game, for example, they will subscribe to a wait list or whatever, or just watch a video that you describe your idea, and you measure how people are, committed to this idea. and then, Lovable came. So a- as you said, it was super easy and, not so expensive to build, like, a real application and even measure how people use the application, not only if they commit to pay for it. and then, you mentioned also that, enterprises and big companies are afraid of such, vibe coding stuff because it's a little bit like a black box.

[33:04] They don't own the, the whole infrastructure and everything. So, this was kind of a blocker, from my point of view, trying to convince big corporation that they n- when they do innovation, they need to validate their ideas before that, and they have the tools to do it, but, actually they couldn't use such tools. So it's very interesting how you advanced this, in your story. And, I'm very interested to understand more about the use cases that, go through this. If it's not like, I don't know, secrets or whatever you can share, it could be super interesting for the listeners. So, yeah.

[33:40] **Vladimir Petrov** my experience with our clients is that, because they are, e-especially one of them is a fairly big, in-international corporation, American corporation, they are very risk-averse. they don't like to experiment that much with new things, so you have to kind of prove the value. So we've been introducing AI behind the scenes more or less.

[34:05] and we keep talking about it. we avoid trying to-- we try to avoid talking about how this is going to affect people's jobs because people are mostly afraid that this is gonna take their jobs. I have a take on that. I don't think that this is going to happen. I think that the jobs will change, and that people will still, figure out how to be employed and how to bring value.

[34:27] It's just that the pot will grow, you know, because there is so many things

[34:31] **Danail Arapkuliev** need to pay the token, so

[34:34] **Vladimir Petrov** True. True. Well, although they are getting s-- they're getting cheaper, per token, that is. So, a side note, he's saying that we should be measuring intelligence per token, and if you measure intelligence per token, it's actually getting a lot, lot cheaper right now.

[34:49] Because the things you can do with a million tokens now and the things you could do six months ago, there's probably a, you know, 10 times more that you can do now. but yeah, I mean, there is an explosion of builders and people building stuff and vibe coding stuff, and, enterprises are, in-- at least in my experience, they are terrified of this.

[35:12] some people inside, they're terrified because of the, absolute neglect to good software practices. just vibe coded stuff is everywhere, and most of it is terrible. It's really horrible. It breaks down, it leaks information, it's just so... You know, they don't want this. they are trying to stay away from...

[35:32] So if you talk about vibe coding, in the enterprises, people start getting really, really, you know, stressed about it. That's why I don't think that Lovable can go into enterprises because they have this aura of quick and easy, and quick and dirty. So we need a product that is enterprise-friendly and that allows, businesses to adopt them, large enterprises to adopt it and to prototype ideas quickly, as you were saying.

[36:00] I feel like this is still an unmet need. and none of them are gonna use Lovable or, what was the other one? There is many, Base44 and,

[36:12] Zero... Yeah, Replit.

[36:14] **Vitaliy Filipov** 

[36:14] **Danail Arapkuliev** But, you know, sometimes I'm also thinking maybe they somehow tried it, selling to enterprises, but they got the same objections as you did, maybe more massively. And they said, "Okay, maybe it's easier for now to introduce do- this through, like, startup owners or, like, really solopreneurs that can, are not dependent on such stuff and are not afraid to try it." but back i- in their strategy, they, I think they have some ideas on how to, like, how to implicate this in the enterprise sector.

[36:52] **Vladimir Petrov** the money is in the enterprises. It's not in,

[36:55] **Danail Arapkuliev** Exactly.

[36:56] **Vladimir Petrov** you know, random person that will create a, I don't know, a calendar app. It's not,

[37:02] **Danail Arapkuliev** is better marketing because, everybody is very, like they love that what they did because it's so amazing. They never had the chance to do it before, and they are starting to to show it everywhere, and this creates, network effect. So

[37:18] **Vladimir Petrov** Well, my vision is to combine these, with a, with an agentic factory or software factory and be able to, iterate on the product that you're building inside this enterprise, whatever it is, iterate on the product faster. So, and maybe build versions. So instead of going through prototypes or HTML pages or PowerPoints or whatever people use to create prototypes, and they get signed off, from the business and then trickle down all this to the developer- development teams and create, you know, all the extra artifacts.

[37:55] Perhaps we can simply create a version of the product that has the, you know, the features that they're asking for and just deliver it to them. Just say, "Okay, is this something you want?" If it's something they don't want, you throw it away 'cause it's all code. But you start from the code base that already- it's already there, and you build something that already works inside the product.

[38:16] And I feel like, Agents give this ability to even non-technical people. I know this is like, you know, why would he say that? You know, developers will go, "Oh, this guy's crazy." You know. But, they are so good, agentic, you know, coding agents, LLMs at the moment, especially if you give them the proper guidelines and you put them in an environment where it knows what, what it should build and how it should build it.

[38:42] you can give it to non-technical people and say, "Okay, you know, as... Talk with the business. Whatever they want, you can build it here." You can show it to them. once they're happy, it goes into a development team. Development team looks at this, refactors whatever they don't like, apply the best practices, do the security testing, you know, do the penetration testing, do everything else that we do, and then just deliver it.

[39:06] So I think it's super, super possible to do that

[39:10] I'm not sure what this will bring, to be honest. it brings, again, you need to end up with someone that owns the product to say, what we want and what we don't want. Because then if everybody keeps adding-

[39:23] **Vitaliy Filipov** this reminds me of when, when we first met, you said a very funny joke about the last man. So it was something like,

[39:30] **Vladimir Petrov** Yeah, what the...

[39:31] **Vitaliy Filipov** factory.

[39:33] **Vladimir Petrov** Yeah.

[39:33] **Vitaliy Filipov** the joke? Can you

[39:34] **Vladimir Petrov** Yes. What is the factory of the future going to look like?

[39:37] will be fully automated, and there will be one person there with a dog, and his only job would be to f- feed the dog.

[39:44] **Vitaliy Filipov** 

[39:44] and in this, in this, regards, what do you think will be-- what will happen after the last engineer? If this man is, let's say, the, in the last engineer in the future factory

[39:54] **Vladimir Petrov** Well, you can take this idea to the extreme, obviously. even if we're talking about AI, not only about engineering. AI is, such a technology that it can affect us to an extent that we cannot even recognize right now. Because if AI can build robots, then the robots can build other robots, and the robots can, work in the mines and to extract all the materials that we need, refine them, do the parts we need, build the power, power plants that we need in order to power all this, 'cause let's face it's, you know, we're lacking the energy for this.

[40:36] So when you put this to the extreme, we end up with a society where very few people do anything us- useful because they just... Anything you want may be given to you by the AI or the robot or whatever, you know. the economists will say that this is bullshit because there is limited resources, obviously.

[40:56] You can't-- all of us can't have, all eight billion of us, we cannot have, villas in Monte Carlo, for example. Yeah, there is nowhere to put all those people. and you can't certainly have so many planes so that you can take everyone everywhere all the time. So there are some limits, but I see a future where a lot of our basic needs are just covered by a lot of automated processes.

[41:23] And when it comes to software engineers, I hope that we have a place in this, but at a, at a, at a spot where creates the most value. I think that software engineers need to go back to basics and, figure out how they can create value, because value is what matters. It's not the code, it's not the what web app I built, or look how I vibe coded this website or any of that.

[41:46] it's the value that matters, and this is how this will play out in my head. So we need to figure out where the value is.

[41:53] **Danail Arapkuliev** Mm-hmm. And what, what's, what's next? I mean, you mentioned at the beginning it started with language, then, the next best thing was to automate code writing,

[42:05] **Vladimir Petrov** 

[42:05] **Danail Arapkuliev** So what will be the next, like automation? And we do another show, and call it after the last, will be that word?

[42:15] **Vladimir Petrov** I think that the web is dying. It will die in the next 10, five years, five to 10 years. we don't need websites. How many websites do you use? Just

[42:26] **Danail Arapkuliev** Well, I still u- still use some, but, to be honest, very often I refer to the agent to tell me stuff and, yeah. On- only if I want to see some pictures and, go a little bit closer to the atmosphere that the agent created for me, replicating the ideas from those websites. But yeah,

[42:45] **Vitaliy Filipov** I'm doom scrolling social, by the way. I'm not sure

[42:48] kill that.

[42:49] **Vladimir Petrov** I think games will, will remain. Doom scrolling for sure will remain. People like looking at other people and their funny, sexy, whatever, you know, stupid thing they do. but actual websites as sources of information and things you can do with them and web applications, I think they will die because even now agents can build UI on the fly for you.

[43:15] you know, yesterday I asked my Cursor agent, you know, "Can you, can you look at our environments?" Because we manage several... One of the projects that we manage, it has environments in Australia and New Zealand, all across. So instead of clicking through portal, you know, in Azure, I went and like, "Can you give me, you know, tell me what is the memory and, usage and CPU usage and how-- what is the load and all of this?

[43:40] Give me the metrics." And you can do that easily in Azure. You can build a dashboard, you can write a script, whatever. But I'm lazy, so I asked my agent, and it built a dashboard for me. And I'm not sure-- I'm sure this is not news to anyone. You know, ChatGPT builds dashboards all the time. You don't need to go into, your bank, mobile app anymore.

[44:03] Why would you need that if the agent can access it and can just tell you what you need to know or act on your behalf? You don't need to go into the website of your bank, or you don't need to go to the website of your news agents or w-whatever, you know, whatever other website we do. It-- They should be able to...

[44:23] and now what they're doing is they use, web automation and computer use, because there is not enough infrastructure, there is not enough n- there aren't established protocols for agents to talk to each other and to, you know, talk to the existing, systems of record. But as those systems of record get more agentic, you end up with agents talking to agents and discussing things agreeing things.

[44:48] and for example, if you want to get a flight booked, you will tell your agent, which will talk to the hundred or so agents of the other, of the, airline companies, and will get you the flight. Or if you need a hotel, the same. You don't need to go to booking.com. If you need... So, you know. So the way I see it is, I think that the web is like a crutch because we cannot deal with the binary information, so we need nice pictures and all of this.

[45:17] but when the agent is talking to other agents, they don't need nice pictures. So they just need textual information. That's it.

[45:24] **Danail Arapkuliev** 

[45:24] **Vladimir Petrov** they can build the nice pictures for you, no problem.

[45:27] **Vitaliy Filipov** All those years I've spent my time, centering a div are now wasted and

[45:33] CSS.

[45:34] **Vladimir Petrov** yeah. I have, I still struggle to do it vertically

[45:37] **Vitaliy Filipov** Yeah,

[45:38] **Vladimir Petrov** to send to it vertically

[45:40] **Vitaliy Filipov** yeah, yeah. This is crazy. And, I'm curious maybe to talk a little bit, bit about the build versus buy,

[45:48] **Vladimir Petrov** Yeah

[45:48] **Vitaliy Filipov** dilemma that we are usually facing, especially in enterprise, because in our consultancy work, this is always the case. Like some client or some, established business, they just want to go digital.

[46:01] They want to do some innovation. They want to build a product, whatnot, and usually this type of question always pops up. So what do you think about this dilemma, and how does it look for you

[46:15] **Vladimir Petrov** Well, it depends on the organization, yeah? So some organizations would-- they want to own the technology end to end. One of our clients is like that. They are like, we don't want you to, offer us ready solutions. We want this built, and we want it." And they have all-- They've bought all sorts of companies that, are not really doing anything, related.

[46:36] I mean, they're do- obviously doing something related to their business, but it's not their core business. It's something peripheral that they have to use, so they just buy the company and in-integrate the technology. Yeah. So they think that they should own, yeah.

[46:52] **Vitaliy Filipov** 

[46:52] **Vladimir Petrov** we had a few projects with them where we basically were-- they were firing their vendors.

[46:58] so they would fire, for example, Tableau. You know, Tableau is this huge dashboard, whatever software

[47:04] **Vitaliy Filipov** Power BI alternative, I think.

[47:07] **Vladimir Petrov** Yeah. So We replaced it with Power BI, funny enough, because they were trying to standardize on the Microsoft stack, and I think there is another project coming up that we'll need to replace it with something else.

[47:18] I don't know. It's, as I said, they w- they want to own the technology. and there's obviously, some companies that will, like to simply compose other, products. They will compose technologies from different vendors, and they will use that to run their, their business. So they will... I think this is actually a good approach, especially in the, in the age of AI because, switching cost is not so high anymore.

[47:49] there is so many vendors for so many things right now that you can experiment, you can try it out. It's not a big deal

[47:55] **Danail Arapkuliev** It's very different from before, yeah.

[47:58] **Vladimir Petrov** yeah. Yeah, before you would have, before if you build it was a form of, a competitive advantage. We have this technology, you know, we are doing this this way. now everybody can build it.

[48:10] So,

[48:11] **Danail Arapkuliev** we have actually partners that are almost not, not buying any, like, any service tool or whatever. just

[48:21] **Vladimir Petrov** it themselves

[48:22] **Danail Arapkuliev** something on the fly, and it's 10 times better and a lot more cheaper than whatever they could buy on the market. So yeah

[48:32] **Vladimir Petrov** Well, there is the argument that if you build it for yourself and solve your problem, this gives you advantage because it just solves- you don't have to learn to use something external that may maybe go away tomorrow, or they may raise prices or, you know, whatever. Well-

[48:51] **Danail Arapkuliev** Hmm.

[48:52] **Vitaliy Filipov** Yeah. D-d-death to the generic SaaS product.

[48:55] **Vladimir Petrov** Yeah. But at the same time, you know, when you buy something off the shelf, then you're, you're also buying, you know, somebody else's, like experience to build this. They're not serving just you, they're serving any, all-- They've seen all the problems that you might have, so they've solved them already.

[49:14] you're b-b-buying something with the, idea that this thing is scalable and it will, you know, it will, it can handle your... You don't have to worry about scalability. It's somebody else's problem. So I think that it depends on the organization again. Some organizations are happy to have everything in-house and then have to worry with infrastructure and scalability and all this, and maintenance.

[49:37] whereas other organizations would just outsource this and say, "Okay, this is good enough. We buy it. They have to worry with keeping the lights on. we just have to use it, and that's it." So super, super different between organizations, I'd say.

[49:53] **Vitaliy Filipov** and what about Cobalt, by the way? Do you think, people should their own software factory or rent something existing that is already the market?

[50:03] **Vladimir Petrov** well, I don't wanna tell what people should do. I feel like there is a segment of companies out there that they want to simply configure an existing solution to their own process, and that's m-my idea behind Cobalt because it's not only a software factory, it's also the sandboxing, you know, orchestrator and infrastructure underneath.

[50:25] So it's vertically integrated, which means that you just work with one product, it does everything you need for your software team. but also it is flexible enough so you can configure it because I've explicitly, left it at the level of pro-- of primitives. So, you know, a primitive is a, is a ticket, a primitive is, a task, which is like a, you know, a cloud.

[50:47] an agentic task is basically we spin up a computer in the cloud, and we put an agent into it. So you can use the task to do anything. I use tasks to orchestrate my other agents, for example. I use it as, for anything. For example, I use it for, cleanup. I noticed my, server like, space was going down because we create so many tasks that it kind of pulls up.

[51:11] So I went, opened another task, and I said, "Just look through all the tasks that are already merged and, you know, everything, and just clean them up." And it just did that, which gave me an idea to do this automatically. So maybe the next thing we'll do in Cobalt is to actually do this automatically.

[51:27] But, the point is that you can adjust it to your, to your workflow, to how you want this set up. You can give it as much or as little, power as you want to. So if you want it to do a lot, you can make it do a lot. If you want it to only handle simple tickets, you can do that. we currently are using it even for feature development because you can, you know, get a task running and, the task, runs the whole application, including databases, you know, s- like everything, all the dependencies inside the VM.

[52:01] And then you can have several. And when you have several, you can work on several tasks at the same time. you can, obviously, you know, you don't have to wait until one thing is done, and you end up with those environments where you can experiment with stuff So I'm exploring this because obviously this is a different use case.

[52:19] This is a use case because it's heavier. For a proper development environment, you need a, you know, proper machine. So you need a VM that has 32 gigs of memory, for example. You need, you know, 16 CPUs or things like that. So this puts some strain on our infrastructure, but it's an interesting use case where you can remotely log in and just everything builds and runs, and then you click a button, and then you get another copy of that, and then you can iterate on this

[52:47] **Danail Arapkuliev** Like how, how people can actually try Cobalt. Maybe you can just explain in a few words. I'm sure that listeners may be interested to

[52:55] it a try and

[52:57] **Vladimir Petrov** yeah, it-it's live. It has a free, free plan. we are currently giving away 300 credits, which is, I think, how much is it? It's about 300 hours of compute, I think. it depends what type of compute you use, so it can be more or, or less than that. So you can use it for whatever you want and ex-experiment.

[53:18] And if you want to, if you want to build something open source, I can obviously give you more credits. So if you wanna bring your open source projects to it, would be great. Right now, the idea is that we need to put it into as many developers' hands as possible because I want to see, where this takes us.

[53:37] I wanna see what are the use cases, how we can make it better, how we can improve it, and how we can, you know, how our, project, products, and processes will change when you incorporate this. So it's super early, I think, in the industry right now. Everybody's experimenting with software factories, but, I don't think anyone has the right setup now.

[53:59] So the more we play with it, the more... the better we'll get. We'll come up, hopefully, with a, you know, with a, with a form factor that, yeah,

[54:09] **Danail Arapkuliev** Cool.

[54:10] **Vladimir Petrov** optimizes

[54:10] **Danail Arapkuliev** Thanks, thanks a lot for that. I'm sure people will, will try it out. We did, so yeah.

[54:16] **Vladimir Petrov** Yeah, I saw your register

[54:18] **Danail Arapkuliev** listeners will do it as well. Thank you, Vadi, for everything.

[54:22] **Vladimir Petrov** Yeah, no problem. Thanks guys

[54:24] **Vitaliy Filipov** this was a real pleasure. Thank you indeed

[54:27] **Vladimir Petrov** All right