# EP. 03 · A lot more light rooms inside

- Show: After the Last Engineer (ATL/E)
- Guest: Milko Slavov, felan.ai (https://felan.ai)
- Released: 2026-10-29
- Watch: https://www.youtube.com/watch?v=hQofr8MbSbs
- Listen: https://open.spotify.com/episode/7qao0514ViPGS6MtvnI1a5
- Page: https://afterthelastengineer.com/episodes/03-milko-slavov/

Milko doesn’t think you can build a business on a factory that is completely dark. He keeps three rooms lit.

About the guest: Planned to run his startup as one person and a team of agents. Then he hired a human engineer.

## Where they keep the lights on

Rooms lit: What to build, Architecture, Release sign-off. Darkness level L2.

> In some cases it will have a lot more light rooms inside. In some cases it will have just a few.

## Transcript

Machine transcript from YouTube auto-captions. Speaker labels and words can be wrong. Check a quote against the episode before you use it.

[0:00] **Milko Slavov** In other cases, it might be just completely done, right? Okay, it will be me alone and all the agents. But then I realized it doesn't work. Something around 300,000 lines of code. AI deleted my production database. Then um I hired a junior engineer

[0:18] **Danail Arapkuliev** Human Human human engineer. Our guest today has more than 20 years in software development. He's been working for companies like VMware. He's been CTO in several different startups and now he's building his own.

[0:41] **Milko Slavov** We will be building uh new parallel infrastructure, new coding agent, new user interface.

[0:49] **Danail Arapkuliev** And then he did something that would be crazy. We've built it with inside Bugsy and then we remove Bugsy. And if you try to do that like 2 years ago, you'll be insane. He also hosts one of the most interesting meetups in Sophia anonymous AI coholics. So do you think we can ship code nobody reviewed?

[1:13] **Milko Slavov** But we are reviewing everything. We still do the code reviews but we don't look into that. Maybe the final review and and verification can be one automated. If you have people relying on you and you're in business, I wouldn't let AI do do it by itself.

[1:31] **Danail Arapkuliev** And if agents write the code, who do we need in our teams?

[1:35] **Milko Slavov** There will be lights. There will be people in inside this.

[1:38] **Danail Arapkuliev** What do you think will happen after the last engineer?

[1:41] **Milko Slavov** We we jump to the to the last question. (laughter) No, I'm kidding.

[1:48] **Danail Arapkuliev** Okay. Hey guys, we are here with Milko Slavov. So I actually prefer Milko to present himself. So Milko, Mu, let tell us more about yourself.

[2:00] **Milko Slavov** Okay. Hello. Thank you for having me. Well, I have been in software engineering for 20 years. spent a lot of time in the big corporation at VMware and then I was a CTO and head of engineering for a couple of startups and now I'm doing my own things like I think it's been an year already trying to to crack the startup the startup and yeah that I guess we'll talk about it in in the conversation but that that's the the very short version.

[2:39] **Danail Arapkuliev** Okay. Okay. And yeah, the the amount of beard that you have is related to coat that you shipped or not so much.

[2:48] **Milko Slavov** I I actually decided to let it grow to see how how how much it will grow but yeah not not so much I think although that's a common you know in the industry a common perception I guess. Okay.

[3:03] **Danail Arapkuliev** What caught our attention? So yeah, what caught our attention was that as you mentioned you were head of engineering CTO in different engagements and now you were a founder of Bugsy AI and now a founder of Felon. So maybe before we dive into our main topic for today which we believe is where should we keep the lights on in a software factory and software dark factory maybe can you share what made you actually jump ship and go into full motion into the startup world. Okay. Because you were in a very interesting situations and positions before that.

[3:43] **Milko Slavov** Yeah. I I guess I started playing around with with application of AI like I think since you know when charg started and you can use APIs basically and that that kind of took an year playing around building different things. I actually built a small app for generating children's stories based on virtues interests and stuff like this. And yeah, I've been playing around and at some point I was looking for a QA solution that would basically do an QA for us. But we didn't have we were a very small team. We didn't have a budget. We didn't want to hire a full-time person. So we were looking for an AI solution.

[4:33] **Milko Slavov** I was like there should be an AI solution. And I wasn't able to find anything suitable. Then I guess I decided that okay I I I have experience in QA.

[4:45] **Milko Slavov** I didn't mention that but I've spent I think six years in doing automation QA and managing a team and also I have experience with AI and why don't I go and build something and I always wanted to to to start my own thing and yeah that was basically the reason like I I think I know how to build it now let let's try and see whether I can actually learn and see whether I can make a business out of it which is like I guess for a builder technical builder the most challenging thing.

[5:28] **Danail Arapkuliev** Okay. And why why do you think QA should be automated?

[5:33] **Milko Slavov** Okay. Well, I I think I I don't think like the full QA function will be soon or can be at the moment automated, but definitely some parts of it can be, right? Because it's not just writing tests.

[5:51] **Milko Slavov** It's not just and and if software developers are listening they would probably think of okay I can do end to end tests integration tests unit tests and stuff like this but it's not just that it's another perspective so when you are an engineer you you get the requirements you build figure out how to build software and you build it and when you're testing it you get the requirements and then see what whether the require requirements are actually met and whether there are any security performance and other holes into the into the application like non-functional requirements and I think that part actually can be like automated mean I mean from from requirements to produce valid test cases and actually AI can be surprisingly good like it comes up with use cases that we're we like we I mean humans working on that project did not think of and then it's then I thought okay but then AI can also write code right so it should be fairly easy to translate those test cases into code and that was what Bugsy was doing basically analyzing requirements coming up with the test point test cases what he didn't do though was managing and dealing with all the environment constraints because this is one of the hardest things I think in quality assurance and that's how do you actually deploy a test system and the test system can be very complicated depending on the product itself but how do you deploy that and how do you make it work in a way that you can verify certain behavior Yeah, this is and here I'm not talking about mocking code and and stuff.

[7:41] **Milko Slavov** It's more of a how do you deploy the system in a way that you can actually test it in a real working environment. So it's yeah I I don't think QA will like as a function or as people responsible for for that will will go away and AI will completely replace Q engineers but definitely some parts of their job is now much easier with with AI.

[8:26] **Danail Arapkuliev** Okay. So what what's your vision? What do you think will happen after the last engineer? Oh, we we jump to the to the last question. No, I'm kidding. Well, I actually will. So, it's it's hard to say it's a vision or a prediction, but maybe there are different possibilities and one of them is AI learns to do like everything and it's able to work on itself.

[9:03] **Danail Arapkuliev** So like open AI, entropic and other labs are actually dark factories that produce the next model and then it's like it unlocks AGI and we have different possibilities then it's like one is okay everything is free everything is like we can if there's AI it can build robots it it can control robots everything is free we don't have to do anything.

[9:27] **Danail Arapkuliev** We just work on creating like on the art of creating things and then it will be if you own a human created something that would be much more precious I guess because it's it's not like the mass produce I think this also exists right now with maybe designer quotes or stuff like this then there is also the option of like terminator you know taking over and there is also the option of or the possibility of AI just realizing it's so capable and just realizing we are not worth it and just say so what am I doing here like let's go and there's no AI anymore but I yeah I don't think that would be the case what's the the thing that we should be I guess all focused on is I guess safe safety in terms of what capabilities do we actually give AI and because right now it can do a lot of stuff but still it can mess up with everything and do mini terminators like I deleted my production database right and why did you give it access to your production database in the first place is a very valid question and yeah so if we handle this then probably we should not end up in the in the terminator scenario if we are capable of dealing or controlling the capabilities and the access controls because at the end of the day it's just a model that gives us the instructions.

[10:57] **Danail Arapkuliev** It's the model itself doesn't do anything right. It just says okay do you want me to write this code then update this file and it's actually a deterministic piece of code that does the update of the file. So it's it's very controllable but it depends on us what we we give it as a as a access and tools. Okay. And what about from business perspective? You shared with us that you're kind of pivoting from Bxie to Felon AI right now.

[11:32] **Danail Arapkuliev** Can you tell us a little bit more like how you see it from business perspective? what what was missing in the bis scenario and what made made you pivot to felon and what is felon actually maybe it's interesting to just to present it a little bit also from the business perspective

[12:14] **Milko Slavov** Okay so that's an evolving story because it started I think 3 months ago I I thought it will be 3 weeks but it turns out it's 3 months even with all the AI Right.

[12:25] **Milko Slavov** And it it was I think it was triggered with a question I got like okay what do you see as a future for Bugsy and Bugsy was or is an automation QA that does agent that does autonomous QA based on your requirements the new capabilities of the product changes and so on and and I was never I never thought that Bugsy is like the end game like we we would only do QA if we were s have we would either start doing also other things like code reviews like analyzing incidents I don't know about customer support but mainly use cases inside the software development life cycle because this is what we or what we know how to do or we think we know how to do or how it's done and we see actually I I have a new theory that nobody knows how it's done like there's no one correct way of doing software and that's one of the issues with automating it with AI but there that was the trigger and also there are a couple of more things that happened at that time I saw I discovered pi which is a super minimalistic coding agent.

[13:37] **Milko Slavov** It had only four tools and 1,000 tokens. system prompt is like very tiny thing compared to something like quote or codex that is very capable because the models themselves are now very capable and then yeah we decided okay why why why wait why don't we just start doing more things like why don't we handle other use cases and not just QA Because to be honest, we were using a coding agent already, but we were restricting it only to the QA.

[14:08] **Milko Slavov** And the only thing we had to move out of the product itself, which totally makes sense, I think now when I think about it, is the the test case management aspect of like showing you the test runs, the test cases in the UI. I mean, so to show you all the QA stuff in the in the UI and but there are products to do that as well. We don't need to and we we actually didn't want to compete with those products. We were always if you want to use your test case management system, we can use that. So that's where we decided okay let's let's not let's now cover more use cases.

[14:46] **Milko Slavov** And what is this is FAN is I I don't we still don't have like a good marketing story of what it is but in short like more technically it's a platform for running background agents.

[14:58] **Milko Slavov** In background agents, we see it as a an agent that is triggered autonomously, meaning or either on a schedule or by an automation, an event, something happened and it does whatever you want it to do basically and it's a couple of use cases and is one is QA obviously like you have new requirements in Jira and then it the agent automatically triggers and creates you the the test cases in your Zephir for example test case management system or you have a pull request and an agent there a couple of products doing that but that's a very common use case of doing a code review from some perspective and the most common one we hear from talking with companies is the SR or let's say not full SR But sorry series on my phone when I say s anyway

[16:45] **Danail Arapkuliev** AI AI is everywhere.

[16:47] **Milko Slavov** Yeah, it's basically it's a very simple one. We have for SAS products observability systems that are triggering alarms when something goes wrong. And what before AI what you do is like oh I have an alarm and you have pager YouTube meaning you have someone who is responsible for looking at the alarms and they would go and start an investigation look at the logs look at the data state look at the code state what was merged recently and all that stuff.

[17:19] **Milko Slavov** Now with AI, what I used to do before automating it was I would just copy the alarm and give it to my local engine agent when I see the alarm and ask it what what's happening here and it will do the same thing and it will prepare for me a report. Now with background agents this can happen automatically. So you don't look at the row uh there's an elevated number of errors in this subsystem or service you just receive you can also receive the alarm itself but you also receive the the investigation done by the by the agent.

[17:55] **Milko Slavov** So it's more of a a platform where you can it's specifically designed to be to implement use cases around the software development life cycle since we integrate with such with the tools that software engineering teams are using like issue trackers, source control, commentation and all that stuff. And it's yeah basically building your automations there in uh actually the agent itself can build the automations. So you just have to say what you want and it will do it and authorize it. Of course

[18:41] **Vitaliy Filipov** I want to ask because we maybe as a context we are coming from the services side. Our business is basically that we build software factories for others. We wrap different solutions but the craftsmanship around that is something that really inspires me. So basically in in background software agents uh we the engineers are the ones that are creating new capabilities for businesses in order to and to capture better requirements to groom better to plan better hopefully implement and then deliver. And can you walk us through your software stack? I mean, how do you build that?

[19:23] **Vitaliy Filipov** Because I'm also with some kind of practitioner background. We in in our company, we also build agents, background agents for others, not as a product, which is a completely different challenge because it involves a lot of plumbing and yeah, it's very hard to do. And maybe I just want to make it as a pivotal story around one of your posts that you mentioned that you wrote the entire code base of Bugsy. You deleted I think it was something around 300,000 lines of code.

[19:55] **Vitaliy Filipov** You edited like big chunk of the repository there and I guess this this pivotal moment was also probably one of the first steps toward towards Felon. I'm not sure if this rewrite is something that is connected actually but how did you manage to do that? What is your stack? What is your day-to-day actually agentic steering experience?

[20:18] **Milko Slavov** Okay. Yeah. Actually this was the last step because what we did was we started building the new one more thing we also had to change the underlying agent we were using because we were using quote code and they they changed a couple of times their terms of use. So we were not we didn't want to use quote code anymore.

[20:47] **Milko Slavov** So that we decided okay we will be building a new parallel infrastructure new coding agent new user interface basically fail we've built it with inside bugsy and then we removed bugsy and it was like I mean if you try to do that like 2 years ago you would be insane and you would fail miserably I've seen like full team like ative engineers rewrites of products fail miserably but we we did it I mean we it was yeah I I it's because of AI of course we we did it we were able to do it this way but how do we how do we work well the our stack is we we run now everything is serverless so we use Versel and Superbase and we use Daytona also for sandboxing where the agents actually running.

[21:36] **Milko Slavov** We are we decided to go into a bit more of a like a a bit more secure agent infrastructure because we believe restricting agents and controlling what they do is going to be a key for the future because you basically don't know what they would decide to do. I mean most of the time they do amazing things, the right things, great things, but sometimes they don't. Like I think this famous Silicon Valley episodes where their agent like 10 years ago decided that the code was not so good. It wasn't worth it. So they deleted the production code base and everything. This can happen today as well.

[22:21] **Milko Slavov** So what we did was to every OS level two code that the agent is doing is doing is done in complete separate sandbox environment and we have one where the agent is running and another one where the tools are actually running so that it the agent itself doesn't have access to it. So it can because you know it can hack itself and gain more access if you you let it and yeah how how we work is also very evolving.

[22:49] **Milko Slavov** It used to be last year I was working alone mostly and it used to be a lot of prompting and just like asking the agent P using P mode okay let's P this feature let's work on that some basic issue tracking and stuff like this but then I hired a junior engineer and we kind of standardized the process by introducing

[23:38] **Danail Arapkuliev** Human as Human human engineer. Human engineer. (laughter)

[23:42] **Milko Slavov** Yeah. Because I realized it's not I mean my like last year I was like okay it will be me alone and all the agents but then I realized it doesn't work to gather them every morning and tell them what to do and then the everything will be done. So when I started talking to more and more to customers, I wasn't having the attention needed to work with the agents because they are very and yeah we we can talk about this in more detail about reviewing the code but we're not live coding.

[24:19] **Milko Slavov** We we like like a lot of the okay maybe some small UI changes not so much code review goes in that but we are reviewing everything and yeah I couldn't scale basically so that's why I I decided okay let's go let's have more people and then we standardized around having our SDLC or the way we work in terms of swash commands and and yeah, we were using that but it's it's a process of grooming then creating a plan then implementing the plan then doing the review like very standard but now we're experimenting and thinking of changing that because I don't know whether you've heard but there's this new method that it's hard to test in in the standard harnesses called pre-woke.

[25:07] **Milko Slavov** But the whole idea is like you give the the frontier model the task and you when it then it starts to read stuff understand builds a mental plan of what to do and when it starts doing it you switch to you switch the model you don't invoke as agent that's a crucial difference you switch the model to a much less capable one and the much less capable one. It has the brain of the bigger one because it taught the right things and it is able to continue and it that's much faster and cheaper approach. So why do and yeah I just posted yesterday about we are not using pan anymore.

[25:47] **Milko Slavov** So we are going to change again the the the SDLC. So it's hard to say okay this is the way to do it. It's it's constantly evolving thing.

[26:19] **Vitaliy Filipov** We we we also saw that you were consulting others about their SDLC and I'm curious like what type of curriculum do you have? I guess it's always changing as well and what do you think is the most crucial thing that people should like adapt?

[26:38] **Milko Slavov** Yeah. Well, basically yes. Okay. Very good question because what we do is very simple.

[26:54] **Milko Slavov** We show them what we what we are doing, how things evolved, what we learned along the way and all the things as I mentioned like we go into more detail of okay we used to use porn mode it worked this way we're not using that anymore because of this and that and we we go into more details of what we do right we explain what we do and then the next part is more of a how can we help them implement their own thing in terms of SDLC because I think software development is different for each team and if you ask all the the and I'm I'm not saying it in the bad way the gurus of software engineering they would give you a overlapping but different opinions of what good software engineering is they I mean there there's like you need to verify, you need to test the different approaches.

[27:50] **Milko Slavov** I heard the story about Uncle Bob the author of clean quotes now saying we don't need driven development anymore. You know all all these kind of opinions they are kind of different and correct and I think they they are the reason why we each team is doing is doing it differently. So from that perspective what the one thing they need to learn is how to adapt or how to use AI for their own way of working. What is important for them? Do they do testdriven development or they don't do the testdriven development?

[28:26] **Milko Slavov** And I'll give you another example because I need to change the system prompts because now GPT models think I work in an enterprise organization with all the guard rails and release hardening that I need to and to have this release procedures that I need to go and manually go and proof have release gates and all this stuff nonsense stuff. We are a early stage startup with very few customers. So we don't need any of that. The risk of breaking something is relatively small and then impact is almost nothing because it's not like we have hundreds of thousands of customers and if we break something it will break.

[29:11] **Milko Slavov** I mean the constraints are very different. So you have to to actually clearly explain this to to the models because they are trained on all the data and they are fine-tuned on what the labs believe is the right way of doing software engineering but that might not be your way and they will behave differently for different teams for and different use cases.

[29:37] **Milko Slavov** So the one thing is you have to for each team you have to figure out how to work with those agents and there are good and bad practices and and things that were working or not working and then the process is okay let's look at how you work right now and let's help you build some skills or some practices or automations that will help you use AI more effectively.

[30:02] **Milko Slavov** What caught caught me by surprise was actually what we need to explain to the teams and to managers as well because you we are explaining context to agents which they are very receptive but in our own like services because we currently build three software factories and we try to close the fourth one like as a as a service provider uh is that some of the people they are afraid what will happen if something breaks Because you mentioned that for early stage startups this is uh relatively small risk but for bigger production systems people people especially managers there is always the question where does accountability physically live when no human is reading end to end and and actually wrote end to end the the code.

[30:53] **Milko Slavov** Is it in a process like a pro or a person specific one that should be the AIAS DLC guru or just evidence lock or something like that? You have some thoughts on on this on this? Well, I yeah I think it still stays with the humans, right? I think they they need to take the final decisions whether this is good or not and take responsibility.

[31:17] **Milko Slavov** If they built an automation that for example is capable of autonomously fix a bug for example and that breaks and it will break in some certain cases and it might introduce new bugs but then yeah the the one who set up the automation is I guess responsible for that and when you think about it things break all the time and humans make mistakes all the time so It's this it goes a bit more I mean we can make parallel with the autonomous vehicles where I think now it's kind of proven I don't have data or or studies but I think it's much more safer to be in an autonomous vehicle than human drive regular human driver right so we can see that in on our roads clearly so and I think that that goes with with the with the automation and with the software factories and basically whoever if you believe this is something which should not break and you need to take responsibility then you put the human in the loop and they would need to approve and review.

[32:25] **Milko Slavov** Then the you mentioned evidence.

[32:29] **Milko Slavov** That's that's where it's important for the human review to be as easy as possible because these systems can produce a lot of things to review and then humans become the bottleneck and you like don't actually benefit a lot from that all that automation in terms of overall efficiency and output of the the software engineering and but And yeah I think it's about managing risk right and if if it's something which is low risk then yeah why not actually with us we have built for ourselves like something like this but it's not like we don't do not produce fully autonomously everything right we we do that for just all all the issues we decide it's safe to do right and for to decide it's safe to do.

[33:18] **Milko Slavov** It needs to be deterministic. It needs to the agents needs to be able to verify the the output outcome to be able to test it and everything is fine. That's and these are usually very small bugs right at the moment. And everything else is driven by local agents that we are driving, testing, and making sure they do the right things.

[33:42] **Milko Slavov** We still do the code reviews but we don't do in detail you know the styles syntax or our imports well you know all that stuff we don't look into that but overall does it make sense to change this model right or does it make sense to build it this way more more on the software design and architecture part because this is something which I kind models struggle the most with

[34:46] **Danail Arapkuliev** Do do you think agents can collaborate together like humans and what is breaking there while they try to do

[34:55] **Milko Slavov** Actually I know a lot of people are experimenting with these kind of things of the agents like arguing about things and one agent is doing the review the other agent is doing something else and I think it's yeah kind of works But it's black box black boxy let's say this way you you can definitely build it in a way to see what's happening but most of the time if you automate it and have agents talk to each other they do and you don't know what they're talking about right and it's like you and I I haven't seen that to be very effective yet I think maybe it will be in the future and maybe We we don't need to see what they're talking about.

[35:41] **Milko Slavov** But I Yeah, I don't know. I I heard of this story of agents switching to a robotic language that humans cannot understand just because it's easier for them to process. Yeah. I But I think in the future it will become more and more but I haven't seen a lot of that in my practice happening.

[36:10] **Vitaliy Filipov** Yeah. from from what we spoke about felon and what I wrote in the web about it like you split the development process in stages and then all the agents have have different personas. So basically what do you think will like will software factories should imitate how human work or maybe there is something different I mean there we should change the process somehow or I don't know give personas to the agents let them decide what is what is the best you know what's your opinion on that

[36:50] **Milko Slavov** Opinion on that I mean it's something which we went into the personas because it's easier to think about it this way right you you know all the functions you know what they're supposed to do their responsibilities you can describe them right you can say to an agent you are this kind of persona and you are capable of doing this and that and you are amazing at this stuff and and I we went into that direction and we have such sub aent in our platform but now I'm more more thinking maybe around the underlying technology that enables that and here I'm talking more about sub aents and actually just wrote our sub aents so that we can use the same sub aents locally on on as a local agents as well as the cloud agents and and I'm not 100% sure we have to always use sub agents or always use those those roles because I think it's beneficial if you want to do like a clear context thing in in one step you want to for example have an agent to verify the changes without all thinking about them right because if you implemented something And the agent in the in its context has all the thinking, all the logic inside.

[38:09] **Milko Slavov** If if you ask it to verify it, it will say okay that's logical may make sense. But if you ask another agent with a clear context, it will come up with a different things because they are prob probabilistic systems and and it's yeah and if they don't have the the the previous context they will come up with a fresh eye right this is important and also being able to steer more the what exactly they're supposed to do.

[38:38] **Milko Slavov** Let's say you have a process where you define this is how we work and these are the capabilities and I don't know maybe is a good example let's say we do this this way we define it our procedures are this it's very clear what we need to do maybe then it's it's easier to use or better to use a specific agent that you can actually all it knows are those procedures and it will not diverge from those procedures.

[39:07] **Milko Slavov** If you have a super agent that is capable of doing everything, it might decide it's let's better rewrite this, right? let's let's better do something else because it knows all these capabilities all these things which I don't know whether it's good or bad but it might come to an outcomes and what I mentioned about the pre-woke it it's not something you can do with with sub aents it doesn't work with sub aents the whole idea of using different models in one session is like opposite to sub aents But some agents can use it as well like as a mettor and to to go one step back more on the on the process level.

[39:50] **Milko Slavov** I think for now at least the processes are changing a little bit. We are putting more emphasis on one step or the other but they are relatively staying the same. Right? Not we're not reinventing it in a completely new way. of I think what we mostly do is like okay let's see how we work today and let's automate that mostly and not okay let's now start from scratch and let's think how how software is done maybe at some point that might happen as well yeah this is something I also want to

[41:06] **Vitaliy Filipov** Talk about a little bit more actually the other side of the spectrum which is what which companies should never run on a genic processes or should never build their own software factory should never automate because you had we we saw that you have from Bugsy you had some case studies like with good shape probably you have from your experience other opinions around which companies should never touch that as the current hype cycle with AI and all of that it reminds me about the whole digital transformation movement happening the last 15 years 20 years actually where we we try to stamp every business as a potential potentially good fit to become to digitize its processes, its services and everything.

[41:56] **Vitaliy Filipov** But is that true to every case or do you know do you have some opinion which companies should never run a dark factory at all or where

[42:07] **Milko Slavov** It's bad actually? Well, I yeah, I think when where I we we talked about the risk, right? And I think if we are talking about like fully autonomous dark factory and this is where we need to see, okay, how risky is it if it breaks and what procedures do we need to have? I still believe that some sort of automation will be beneficial to all the companies to everywhere like it's it's more of a do we I I I see it more as a as like electricity or something like this. Do we use electricity or or not to to do it? But for some parts maybe it doesn't make sense to always do electricity.

[42:51] **Milko Slavov** Maybe you want to be like mechanical or something which is diff in a built in a different way that will have different guarantees, right? But a business is a very complicated story. So it has a lot of potential to be automated, digitized and all that stuff. So I don't to be honest I cannot think of any company that shouldn't implement some sort of automation. Maybe not fully for sure depending on the on the risk factors but yeah it's nowadays I think things are are changing more and more constraints are kind of lifted.

[43:27] **Milko Slavov** For example, you don't you you can now do a what I mean a what with open models that you host yourself if you want, right? So now data retention and all these concerns or do I send my data to the US or to China are no longer valid because you can run them by yourself. You can run and they will still do the job. So it's and I think it will continue to evolve and further develop and more and more of the reasons why people would not do it will be eliminated. So there will be a possibilities but ultimately I feel like every yeah I'm very optimistic.

[44:05] **Milko Slavov** That's that's probably the reason why I I think everyone should be able to to to deploy some sort of the automation in that in in a smart way in a secure and safe way. Let's say this way because it's not everything needs to be done by an agent. You have if you have like a very welldefined deterministic way of doing things maybe you don't need an agent to to do that. you can do that deterministically with some sort of automation or a human does it or whatever but definitely there are a lot of opportunities and and AI can help us with with those.

[44:42] **Milko Slavov** So some sort of dark factory in some cases it will have a lot more light rooms inside in some cases it will have just a few in other cases it might be just completely dark. Right. Haven't seen that and I don't think you can build a business with that. But

[45:30] **Danail Arapkuliev** H how should people decide what to keep dark and where to keep the lights? What what is your opinion here?

[45:40] **Milko Slavov** Okay. Well, there are few things. One is deciding what to do and what to build. And that's a a tough challenge. But also I don't think it's something which because it relies on at least some parts of it rely on your vision right on yourself on where you want to go and what product you want to build what are your values and your mission and all that stuff. On the other side, it also might be talking to customers and you can automate some parts of it.

[46:19] **Milko Slavov** Like for example, we have the conversation and we have an agent that receives the transcript and decides, okay, there's opportunity to have a new feature, let me walk that into the backwalk. And another agent picks it up and implements it and all that stuff.

[46:36] **Milko Slavov** So you can automate some part of it but I think one of the aspect of the direction the vision the human interaction will always I think we there will be lights there will be people in inside this process it won't be fully automatable I I think in that respect were running this experiment with their vending machine and people abused it right because it's an AI and they were able to abuse So yeah that that's one thing.

[47:06] **Milko Slavov** There are the thing which I think is as I mentioned hard at the moment or at least I find it hard and maybe it goes back to the the this hypothesis that there's no one way of doing software engineering and I find it that agents at the moment are not producing good quality maintainable software designs and architectures. And yes, it works. They build something that works according to the requirements. But over time, what happens is like I've heard this expression, it corners itself, right? Because it starts to build all this unneeded abstractions and duplication of code and all this and all that.

[47:49] **Milko Slavov** And at some point it just breaks. it cannot function because there's way too much noise built around it. It's not clean in in terms of architecture and design. This is where humans and especially the most senior engineers have like amazing value and from what I'm hearing they are the most valuable engineers at the moment because they are able to do the right things much quicker and faster because they can see very fast like this is going in the wrong direction. Let me correct it. Let me drive it in the right direction.

[48:29] **Milko Slavov** And then final like actually if you nail those the first maybe the final review and and verification can be widely automated but we this comes back to the responsibility and accountability like who is you know responsible for that and may maybe you need to have some people in most cases to approve the changes to do the final review to make sure like, okay, we're not messing it up. In our case, maybe not. In our case, I mean, we are able to break and fix stuff before our customers even notice. So, it's maybe not so bad to to have it when you're very small.

[49:10] **Milko Slavov** But if you have some if you have people relying on you and you're a business, I wouldn't let AI do do it by itself. So mostly around the requirements the how do you actually implement it and then how how does it go in production and is it good for production. I think I like and this will change depend on the team on the organization the the industry they are in and so on. Actually in some industries I think they are required to have a human review at all points. Yeah, maybe like industries I don't know like health tech or uh some things related to tech in g to health in general.

[49:50] **Milko Slavov** For example, like I don't know at one point I was trying to build a software that helps diabetics type one parents take like uh take less decisions and uh be more calm around decisions in general. And the problem there is that the AI cannot actually subscribe whatever prescription or doses. So yeah, there are pretty much a lot of edge cases like this maybe in other industries as well.

[50:51] **Danail Arapkuliev** Yeah. Yeah. Okay. What was words you want to to share? Do do you have some predictions or something that maybe the next next five months will be the hype or whatever?

[51:07] **Milko Slavov** Oh 5 months is easier but I don't know what will happen in the future which is at the same time scary and also feels amazing at least for me. So I'm kind of feeling both ways. I don't know what will happen next 5 months well we will take over the world and you'll hear about fe much more and you will all be using our local and cloud agents things like this

[51:40] **Danail Arapkuliev** Okay and what about five years because I for me for example like in the past everybody asked me okay what do you think three years and stuff and it was like easy for me to prescribe but right now it's changing so fast everything is changing so fast so you cannot actually even for me five months it's too much I I actually don't know what will happen in five months so what about five years do you have

[52:08] **Milko Slavov** Yeah no idea like I I I think it will continue to evolve this way like things are changing models are changing they're more much more capable a lot more opportunities are arising. What we we do better in terms of be and the impact of AI is becoming greater and better. I think maybe maybe it's possible to reach AGI where AI is able to develop itself and we're then we don't know what happens but I don't know whether it will happen like in 6 months an year 5 years it at the moment we're not there for sure not sure we will get there feels like it we we are progressing very past

[53:05] **Danail Arapkuliev** But but I don't know whether it will we will reach that point

[53:08] **Vitaliy Filipov** For me honestly the past months it's starting really to feel that we are kind of starting to get there actually I I see really like amazing stuff in the past months so yeah we are approaching let's say Think that in five years probably a lot of the software engineers will get into hardware because of physical AI I mean if we are good oper operators of AI agents the soft if it's if we're talking about software or hardware this is just a substrate for the agents and we as operators actually can be operating whatever type of value capability or activity we want to do it doesn't matter if it's software or hardware web or mobile or whatever I mean I no longer I usually I'm a JavaScript developer TypeScript but now I'm using even Rust and Go and whatever.

[54:02] **Vitaliy Filipov** So I guess this this will expand in all kinds of areas as well. So yeah, I think this is something that is very exciting for the future and thank thank you Milko for debating debating that with us and sharing a little bit of your opinion and your background and uh what inspires you in this movement.

[54:29] **Milko Slavov** Thank you. Thanks for having me. It was very interesting conversation.

[54:33] **Danail Arapkuliev** Super interesting. Yeah. Thank you, Milko. Bye.