Why your next hire should be an AI agent with Ivan Burazin

In this episode of Delivered, you can learn how to prepare your business for the next chapter in the AI evolution.

We sat down with Ivan Burazin, co-founder and CEO of Daytona, a fast-growing startup providing a secure infrastructure for running AI-generated code. His vision for Daytona is to make software development more accessible, efficient, and enjoyable for developers of all backgrounds. Ivan previously founded Codeanywhere, one of the first cloud IDEs, and launched Shift, now one of Europe’s top developer conferences. More recently, he served on the executive board of the first Croatian unicorn company, Infobip.

Key learnings:

• Learn what AI agents are, what they’re capable of, and where they fall short
• Find out where AI agents can drive real business value & how to get your business AI agent-ready
• Understand the adoption risks and how to avoid the hype
• Discover what the future holds for AI agents and their impact on society

Thanks for tuning in! If you enjoyed this episode, hit that follow button so you don’t miss our next delivery.

Delivered newsletter
Delivered episodes are recorded live. Subscribe to our newsletter to stay informed about upcoming live events. When you join a live event, you can ask our guests questions and get instant answers, access exclusive content, and participate in giveaways.

About Infinum
Delivered is brought to you by a leading digital product agency, Infinum. We've been consulting, workshopping, and delivering advanced digital solutions since 2005.

Let's stay connected!
LinkedIn
YouTube

Transcript

Hey there. What's up? I just hanging around waiting for you, I guess.

Yeah, what's up Ivan? Welcome to the show. Where do this livestream find you today?

I am. I usually live in an airplane. People ask me where I am. Currently I am in Croatia, so I'm in sunny Split, which is great. And then tomorrow fly it back to San Francisco.

Very cool. So I think we should just get straight to it. Tell us about yourself and your career and how you ended up creating Daytona.

Sure. Career has been a long one. I'll date myself probably with this. So first company was building server rooms, so actual HP, IBM servers, routers, networks, stuff like that. In the early two thousands after that sold, that company did the very first browser-based IDE. So basically in your browser that you can code, which is a very normal thing today. Now you don't even have to type code, it's usually vibe coding. But in 2009 that was very, very new, very first one in the world, so we had to create a lot of infrastructure things on our own primitives because they didn't exist after that. Did a developer conference, which was quite successful. So about 4,000 people attend the conference that got acquired by a company called Infobip, which is this very, very large communications platform as a service, basically an API that enables developers to send SMS, WhatsApp email and whatnot. And after that I left to start Daytona, which is a agent native runtime company. Basically what that means is that we enable AI agents to not only execute code but do various tasks. If you of an agent as a human worker, which we're probably going to talk about today, we are equivalent to their laptop, something like that.

Okay, okay, cool. So it's like a cage where you can put it in and let it do its thing and it'll keep in the cage.

So actually it is basically the agent is running somewhere and then the agent can invoke basically a Daytona, what we call a runtime or a sandbox. And anything done in that is a cage. So you can think about it, you as a human probably have one laptop, but an agent can have a lot of them because it can break them because agents are not that sophisticated. And so you can imagine a younger child giving you a laptop, you'll probably break it at some point or do something. And so in that manner, we can spin these up quite fast. There's more to do there than just make sure it doesn't do anything malicious, although that is one of the themes of that. But there's things like in the sense of an agent, unlike humans, can work on the same thing in parallel multiple times. So if you think have a task, you're going to try to solve it in one way, coding task or mathematical task, whatever it may be, you're going to try to solve it one way.
If it's wrong, we'll do another. If it's wrong, we'll do another. An agent can basically say, oh, this is my machine and I'm going to make it five times and then try five in parallel and say, oh, number two is not bad, number four is not bad, let's take that. Or they're all garbage, let's roll back. Or number one, it's great, let's just take that. So there's more things to that that we offer to AI agents in a sense of also AI agents sometimes need a browser, sometimes they need a desktop to actually test out applications, whatnot. And so we offer that to the agents specifically.

Very cool. So I think we're already jumping into the deep end. So if we back up a little bit and onboard people like myself or people that are newer to the topic and starting with a simpler question, so what are AI agents, how do they work and what's the difference between the LLMs and the generative AI tools that we use today, like the ChatGPTs and Claudes of this world? How would you explain the terminology between these things?

Sure. There's lots of definitions and it's all over the place. And so basically the way I sort of describe it is an agent is a piece of software that on one end is connected to a large language model, so OpenAI, or whatever. And on the other side is connected to tools and the agent itself is the software that connects those two and communicates with the end user. So that, for an example, ChatGPT is a agent, but OpenAI is a model company. So they have the open AI platform. So if you go to anthropic.com, you have access to the model. If you go to Claude, you have access to agent, which is a piece of software that uses their model in there. And so I'm happy to go into more details, but basically the idea is can you communicate with this service, this agent and can it produce, can it solve a problem or an outcome or a question on its own?

Okay. Basically if I would compare that to how I'm using ChatGPTs, like in ChatGPT, I will write a prompt, have it respond to me, and then I will write it the next instruction and then it will feedback and is this that it could go, you give it a task and then it will do the separate steps without interference.

So the way I think about that, so ChatGPT is an agent specifically, so let's try to break that down. So ChatGPT, you go to ChatGPT.com, you see a web service or a mobile app, whatever, maybe.
That interface, that piece of software that is created is the agent. People think that that is the brain, but it actually isn't. The brain is the model, the large language model in this case, which is running somewhere else. So if you think of we can do mobile apps, which make it easier, you install the ChatGPT app on your phone and that is an agent, but the model is not on your phone. It actually has to connect to a model to get the answer. So the way it works on a more technical level is you prompt it, you prompt the agent, not the model. You prompt the agent, the agent takes that data, it's connected or your prompt, it's connected to the model, it's sends it to the model. The model is basically the brain and the model comes back with an answer. So you can say, oh, what is one plus two, whatever it may be, or what's the capital of London?

And if it has that knowledge, it will reply with the knowledge and then you as user will see that, and then you will either be happy with the answer or it will stop. The reason I say an agent invokes tool calling is because that was the very, very first what I would just explained is basically everything that the agent does is basically in the model's brain. So when you work in that fashion, you say, oh, what is the percentage difference between 256 and a hundred thousand? I'm making up numbers. The agent on its own without invoking a tool call will send it to the model. The model will try to solve it in its brain, basically like we would in high school. It will write out the entire formula to get to the answer, which is not efficient because the way we work today is you open up a calculator app and you type in the numbers and you get the outcome.

It's super fast, super easy, super cheap, all those things. And so an agent today basically can talk to a model, but it can talk to tools. So it doesn't have to go to a model for every single solution or even the model mid solution can say, oh, it's like a mathematical equation. I'm using calculator as a simple one. Let's just call a calculator app, type in the numbers, get the answer and then continue on. So very much like a human, as I said earlier, agents, you can think about them as sort of like knowledge workers because a human in today's world is also a brain, a human that is connected to a laptop that's producing some sort of work that someone has input them to do, right? You're a boss, your colleague, your client has sent you a prompt, which is a question and ask whatever request for service, and you have gone at least in the digital world and opened up your laptop and started solving that problem and then coming up with solution. Okay.

Cool. So is there a difference, so I've heard a lot of people throwing around these terms. They say AI agents, and then they describe something like operator by ChatGPT, and I've also heard people saying agentic AI, is that the same word or is that just

Yeah, so basically there's a lot of words. Gen, AI, Gentech, AI, whatever the words that basically AI itself, the people that want to talk about AI are usually referencing large language models. The two anthropic and open AI are the most prominent xai from Twitter and whatnot. There's much more of them. And so if you think of them, the models again on their own are nothing without basically, let's take a step back. When people started thinking about AI is when they got their hands on ChatGPT, all the models existed for chat GPT before we had ChatGPT, it existed, but you as a human could not interface with that model easily. The APIs were there, you could have done it, I could have done it, I didn't. But until someone created an agent, it ended up being OpenAI also did the agent, but it could have been anyone else.
That was the interface to the human. So basically the key thing is the agent interfaces with the human and the model and with the rest of the world. And so it's like a three-prong type of thing where at one end there's a human, one end, there's a model. The other end there's tools, and if you think of operator specifically, so operator for those you don't know, it's service from OpenAI and it can open a web browser and click on things and book your dinner in San Francisco. And so that is a tool. So in the sense of it's the same agent, so it's still ChatGPT is the agent, but what it has access to is a tool and the tool is a browser. And so in this case ChatGPT now has the ability to not just solve problems in its head and write out textual prompts.

It now has access to a browser, a computer with a browser, and then it can go and click on that and Daytona something that we do. We also enable other companies to build things like operator because we offer that computer including a browser for people or sorry for agents to be able to do these things because without it, if you have ChatGPT, without that access, it really can't go out and book a restaurant dinner for you and your partner tonight because it doesn't have access to the rest of the world. It's locked inside of ChatGPT. When you give it a browser, it can do everything that human does. Now does it do it good or bad that we can get into that later, but without having the ability, the tool to be able to do that, it can't actually go and do that.
And the way I like to think about to tell people what tools are for agents is imagine you have your staff and you take everyone's computers away and tell 'em, get their job done. It would be very, very hard in our world if you're, and I'm not saying anything bad, but if you're a farmer, probably you can do that. But if you're a digital agency, if you remove everyone's computers, they're not going to get much done in that same manner. I would get out of a job. Yeah, yeah. In that same manner, agents, if they don't have the tool that machine, they can't do that either.

Yep. Gotcha. Okay, getting clear. So you already touched on it, what it can do and what it's good at and what its limitation. So how autonomous are the AI agents you're working on with Daytona at the moment and what do they excel at and what do they bump their toes with?

So there's a lot of limitations right now. So basically what we'll see in the media is there a lot of hype. It's like, oh, there'll be no more software developers. Oh, all these things are dead. Oh, look at this new demo. We talked about this earlier. Oh, there's a new demo about something. It comes up and demos are demos and so they work really well and sort of like a boxed scenario. And then when you let them out in the wild, they're not that good. Not to say that agents don't have value right now, but they are really, really good at small, short, narrow, repetitive tasks that they obviously can't handle they can interact with and they are very bad at long-term planning at long taking tasks and unstructured problems. So an unstructured problem is usually very hard for humans to solve, let alone agents to solve.
And I'll give you an example. When I was working at my former company, the task was create a meetup in a foreign city that's a task and make it successful. So there are a bunch of unknowns there. So one is what date, what city, what location, how many people, they're all unknown, but the task is to make it good and solve that, right? And so that's hard for the average human because we all, especially in large organizations, a lot of things are either defined or constrained. So they're either defined by the input of the buyer user, whatever, or they're constrained by whatever. So easier task would that be is like, oh, I want a meetup in New York on May 29th in a bar. And so the more I give you of these things, the easier the task is. And so that is what agents are really good at is if you can clearly define exactly what you want, you just want them to go out and do the thing that's amazing and that's what they do really well. But the more obscure the task and the longer it has to run the worst it is, and I'd actually say it can't actually solve any of these tasks at any scale at the moment.

Okay, cool. So can you give some, I mean you just gave a short-term example. What's a longer example that it would have a difficult problem with?

So anything that takes a long time is pretty hard because one, it usually fills up the context window. So it's like, oh, I'll take a coding examples because just easier is if you have a very, very large task that you have to do that will take a normal engineer like days or whatever it was, that's going to be very hard for an agent to do to solve successfully. And just because it has to figure out the whole context of what's going on and the way models work, basically they sort of try to ingest everything if they can and then try to figure things out. And there it usually breaks. I have yet to try a autonomous coding agent. So it's an AI agent that codes and say, oh, here's this not even super complex, here's documentation of our service and now I want you to remove these things, add these things and go off and do that.
That is really, really hard is never been successful. But something that does really, really well is if you have this, even if you have a very large product, think you're working in Ericsson or Bank of America or whatever it is and you're using Twilio for messaging. So we have an API endpoint which is like Twilio's and now you want to switch that out with someone else's, like another service that is something that humans really don't like to do because there's going to be hundreds of places where this is connected, this endpoint that you have to change the credentials, you have to change all these things. Humans really don't like doing that and agents are fantastic at doing that because it's the same exact task repeated multiple times. So when you think about what an agent is good at is, is it something repeatable, short, concrete, you could send it off and do that. But if it's something very large and have to think about and it takes long loops or long iterations to do things, it's going to be really bad to do that.

So if I understand it correctly then, so if you have a really tight brief or it's updating data entry of several different fields in a database, it's good at that. If you can brief it very clearly on what you expect from it bad, if you need to ask it to reason about the problem figuring out and then maybe try different things, bump its to reason why and then try to solve it in another way

Exactly at the moment right now, even when you try such at GBT, again as an agent, we can use that when you give it a very complex thing that you have to do, I doubt it'll actually solve the problem, but what it can do is if you can sub scope some of the elements of the task, it'll go out and do that. And then you as a human have to first verify that what they did is actually good and if it's actually good, then you can sort of incorporate that back into the original content. So I'll give you something very useful I think for most people or very understandable, if you tell an agent, oh, I want to write a blog post, like a long form blog post about whatever solar energy I'm making it up and the more technical, the harder and use deep research, find all this data and write me an interesting article.

You can even say use the style of whatever author you really like that it will create it, but it'll actually be pretty garbage. It's going to be really wide and shallow. It won't go into deep ends. It is not actually really good at that. But what I can tell it to do is like, oh, find me the data of whatever I need about how many solar panels are, how much of the energy is being created, how much energy we need. It will feed these things back to me and then I can basically write out an article which will be crappy, like my grammar's bad, whatever, and then I can pull out sections of that article and say, oh, now that this is the form of the article, can you change the tone without or fix the grammar or whatever without adding it or can you read this back to me and tell me what you understand and what it didn't come off well?
And so those narrow tasks where the answers are really short, those things are very, that's where agents are really useful. And it does save you mountains of time. So if you are going to do a article right now where you have to do all the research, write the article, do the proofreading and all these things, that would take you a really long time, probably a week of your time because you have other things to do. But now if you're doing that, you might do it in a Saturday, maybe a Saturday and Sunday and it'll be like a high quality article that you get out of that. Yeah, figure out the answers. As long as you direct it in the order that it needs to get things done, the quality of the output becomes a lot better. It seems we were keeping it quite theoretical and high level, but it's like what? I mean you're right in it. You're also based in San Francisco. So what's your favorite real world examples of AI agents in the wild right now? What are you currently super excited about?

Well, I'll change the question a bit. I'm more excited about the net new use cases. The example people talk about, and I'll use the same example, is I think there's very few agent native if we can use that term services right now that exist outside of ChatGPT and whatnot. Why I say that is the first mobile native services that could not exist are DoorDash and Uber and whatever these things that needed GPS location and communication at the same time did not exist, could not exist until you actually had mobile. So the agent natives companies are just being built or about to be built or whatnot. And I find that is the most exciting because right now what we're doing is basically porting all old technologies with this new AI flavor basically the same way we ported, it's not exact, but windows for on mobile, for those of you that remember there's Windows C, which it was literally you had the window start button, it was not mobile ready, it was just like windows on a small screen and the web was on a small screen and everything was on a small screen and that was it.

It wasn't utilizing what these things are. And I see most things being created right now or mostly being used just like AI fairy dust, which is like, oh, I have a SaaS product or whatever it is, and can I just add this chat bar and now it's going to reply answers and whatever, which I don't think is actually AI native in the sense of is it actually doing things as it should be? And so some of the things I've heard of and I haven't been using, and some of 'em are customers are well outside of just building coding platforms is so there's an AI native notion competitor. So basically it is like a notion which is a service that you can type things in and it has that data and whatnot in there. And so what they have in the background is they also use us.
Users won't know that. Why they have us is because you can actually tell the notion competitor the product itself. It's like, oh, thank you for this data. Can you now just draw out an analysis or do run some crunch these, you can actually ask it to crunch numbers and then it spins up a Daytona instance in the background to be able to crunch these numbers and gives you the output. It's not a AI chatbot on top of notion. It is trying to fundamentally change that. That might not be the best, but this is what I'm seeing where people are actually trying to find out what are the net new things that can be created now that AI exists and agents are a thing. And what I'm seeing is, so Y Combinator are one of the most famous accelerators in the world. They put out the numbers, 37% of this batch.

So they have about 200 companies per batch, 37% are agents. And I think that's actually quite profound because the way people understand the agents for the most part is ChatGPT, Claude, and that's it. There's no more. That's it. They're the companies in the sense of Google is search and these are agents, but depends on how far we go. And we can talk about this, we talked about this earlier, depends how do we get to singularity or do we not get to singularity? These agents that they have are just like to an extent general purpose, they're not specific tools for specific use case. And I think the way forward and what is happening is that all new software are actually going to be agents. So agents for whatever X they're trying to achieve. And if we get these new software paradigms similar as to what we had mobile paradigms, I think that is where things get really exciting. And I don't think we've seen anything yet at massive scale that is an ai, sorry, an agent native sort of product.

Okay, so it is going to be like you mentioned Uber before where everybody's in venture capital said it's Uber for X. It's like, oh, it's Uber for laundry or it's Uber for food or it's Uber for this.

To a degree, right? It's like not everything is like delivery and having thing coming to you, but basically what the technology enables you to do, because before you maybe had a laptop, some people had laptops, but it didn't have a nonstop internet connection, didn't have a GPS inside of that. So Uber wasn't impossible to create. I just use that because it's the most common one. But there's so many things in mobile that are now used that cannot be used anywhere else or could not be of created if mobile doesn't exists. And so if your product is now an AI agent, how much, what are the things that we can now achieve that we couldn't do with, let's call it dynamic applications, but they weren't agents applications, hence the world agentic. So the applications we have now are not static. They can change, they can do things, but they cannot work for you. They do not provide services themselves.

Gotcha. So next question I have is like, okay, so you got me bought into AI agents. So on this call there's a lot of different people from different types of sectors who should be adopting AI agents and why? I mean I guess the answer is everyone, but which ones are the most relevant to get to in the short to near term? And which sectors do you think should jump on this right now? And are there anything that's just complete hype at the moment that we should?

So one thing I think depending on your audience, and so I'm doing this because is a digital agency that builds digital products. And so I'll add that in here first, which I don't think most people appreciate. I nodded to this earlier, whereas some people should actually build agents. So a lot of people should build agents actually not just use agents in the sense of before. And so I'm assume customers of Infinum and similar companies are like, oh, we need this piece of software and it could be internal, but I believe for the most part it's external. It's like, so we need a mobile app for our brand, we need a web service for whatever we do. I don't know, you're selling tires so you need a web shop or whatever it may be. I'm simplifying it and that is great. And I believe unless we hit singularity that in the near term we will, your customers or customers of infant, more people that are building will actually be building AI agents instead of these credit applications that have been creating historically or credit will still exist, but you will be building agents itself.
So challenge everyone to also think about can I build an agent specifically? Obviously you'll take things off the shelf and examples are a lot of people use Salesforce or similar applications. Very few people try to build Salesforce. The few that I know that I have internally because Salesforce wasn't good enough, just dumped that away and ended up just using it off the shelf. But all these companies do not exist yet. The agent native winners do not exist yet for a specific vertical. So can your company one build the winner in that or two can my whatever may, I mean I'm selling tires, can I create an agent to do X? And so I'm not sure it makes sense for tires, don't know. But what I'm trying to challenge people is you should start thinking about can you create agents to help your company both internally and externally to help your customers?

And I think that is a fundamental thing that people don't think about that you can actually build an agent. There's a lot of tools and a lot of frameworks that allow you to do that. But also there's a lot of companies like Infinum, and I'm not trying to plug you on purpose, but it just makes sense that companies like this can offer not just a web app and a mobile app, but also an agent app just using that word, but an agent itself. So that would be the first thing that I would call on before getting into what agent specifically you should use.

Gotcha. But so then if we continue with a tire example, so tire app or tire website, I have something where I can order new tires or schedule a replacement or a check or whatever, and I want to get into the agentic evolution of my business. What's the measure of success? Is that efficiency or what is it? It depends on what you're interested in. And so if I'm making this up on the fly, to be honest, so people watching this, listening to this, don't hit on me because of that. So if you go into, I was changing my tires recently, that's why I came up. And so if you look at the number of people that are typing in information when I'm changing my tires in the place that I was, there was six people on six computers logging me in, writing this information in, doing this, sending your cars over there, all of that could be an agent. All of that.

Why it's very structured repetitive task. My car is that car, it has that number, it has those tires, that is it. Have I paid or not paid? All of those things could actually be agents, perhaps not to the customer, but certainly for the productivity of the company internally, which gets into the point of will agents replace us or whatnot. But generally as a owner of this business, it is more beneficial to have the people actually changing the tires or changing more tires than having people that have to type in these things which are very repetitive but not perfectly repetitive. So you can't use an RPA because it's not exactly the same because different people come with different cars, different whatever. But an agent is smart enough to do that.

Can you clarify what's the RPA is?

So it's basically automation that people have existed for a long time where you can basically record a set of steps and then the bot can do the same thing over and over again
Where an agent does a better version of that or better work depending on the use case agent can get it wrong, but agent can adapt to different scenarios, right? It's like, oh, what's your car? And you'll say your car, what's your tires? It's like, oh, you already knows the tires because it can look up what tires I had originally or it's in the paperwork or whatnot. So it can do all these things automated and just say, okay, leave your keys, whatever. Some will come in an hour and your car will be done. So what are the things that you actually do not like doing is probably the best place to look at where to use the agent. So what is a job that I have to do every single day that's repetitive, mostly boring, not intellectually intensive. And most of that can today if not completely start to be automated with an agent either off the shelf or being built out.

And what's the stuff that it's not, which I should not let it do so far. So it depends. Risky, depends on how risky you want to be. So all of these things, agents hallucinate. Again, someone said the day humans hallucinate as well, so humans screw up as well. So agents do, but unlike other technologies, it's not deterministic. So in the sense of it won't be the same every single time, even if it has to be the same every single time. So there are things that you still might need people to look over or whatnot. But in the same vein, people that Waymo, the self-driving car company, I've ridden a few times in San Francisco and the way I understand it, it's mostly self-driving, but you still have remote drivers if something messes up. So you don't need one driver per every car, you need one driver per X cars, which is still way more efficient than having that because it's much easier to do that. In the same vein, you'll probably need someone to look through if it's sensitive data, if it's not sensitive. And as we mentioned earlier as well, you ask it to do some deep research, but you probably should if you don't double check the references, it could be wrong, but you're just checking references. It's like, oh, reference numbers good, good, okay, I'm gone. I don't, yeah,

But that's something I would do with a junior employee anyway. Right,

Exactly,

Exactly. Just quick, would you say that Waymo cars are agents,

So I dunno, they're AI and are they agent, so you could define it and you're catching me on the spot one, is there interface with a human? Yes. Two, does it have access to tools, which is a car in the tool? Is the car? Yes. Is it connected to a model? It's a different model than an LLM as far as I know. But a model, yes. So your definition would be that it's an AI agent? Yes, under that definition it would fall. That's cool.

So if we continue, we bought in, I'm just going to go with this tire company reference now since we started.

Yeah, I wish I didn't do the tire company. It should have done something else. Yeah,

It's going downhill. Yeah, it's going

Downhill from here for sure. How can companies like a tire company or any other organization get AI agent-ready? What do I need to put in place or do I just get a company like Infinum or Daytona to like, hey, let's start building agent software and what do I need to have in place for it to be or increase my chance of success?

I'm not sure if this is the right way to think about it, but this is how I think about it. Unlike, so I dunno if you've seen the numbers and people are doing what you're saying. And so I think it was Accenture, one of these big consulting companies made whatever it was, it was like a stupid number, like 10 billion last year on consulting on ai. It was insane. The most money anyone has made it was the consulting companies, although they're having their issues now as well. So I don't think people assume that there is someone that knows better than them, probably in generally, but lemme rephrase that. People assume most other things that you can buy knowledge because what is my best CRM? What is my best ERP? What is my best? Whatever it may be, and just take the best buy that and that's done. But we're basically living in a world where things are just being written so there is no solutions that's like, oh, call this company and they'll get it done because maybe the thing that you need AI is not where it should be and it's useless for you.
The only way I would suggest people is be on top of whatever is happening in the world. Obviously you should use AI in the sense of the general purpose. GenAI tools like ChatGPT, Claude is whatnot, as much as you can for sure and just be on top of how things are progressing. They are progressing very, very fast. But at the same time other things are very slow. What I mean by that is if you live in the bubble of ai, which is mostly in Silicon Valley, which is mostly in San Francisco actually right now, a lot of people are serving the AI universe with ai. So basically we are also a tool for ai. And so the more things that happen in that world, the more we all benefit. But the further you are away, you can see that a lot of these things, the benefits have not trickled down to the tire company. Majority have not trickled down yet,

But things are moving at a very fast speed. So it's like, oh, it's not like, oh, I don't need this, let's forget about AI because it's useless to us. And I've said this historically and I've seen people, it's like they try something, oh, this is garbage, it doesn't do what I want it to do, I'm never using this again. And I think that's a wrong way to do it because things do change really fast and three months it probably will work really well. Why I say this is because one, it's moving, but two, if your competitor and you certainly have a competitor, whatever you're doing is on top of these things. When there is an AI feature, maybe it exists, maybe it doesn't exist now, but when it becomes, if they have that or using it and you not, they will exponentially be better than it will compound how much more they can get done or be more productive than you can and you'll be left sort of in the dust.

And so not an exact answer to your question, but I would suggest, which is hard for the tire owner company person, but definitely should be on top of it as much as they can in the forms that they can. So however they consume information, however they test these things, they should be, not to say they should not call in for them, maybe they can, but the question is for the guitar company, is the time now or is it six, four months from now or is it three months from now or was it already yesterday? They're late? It's hard to tell for a specific vertical where you are, and I say this because for coding, I dunno if you've seen the graph. So big tech, big tech being Google, Facebook, all these large companies, generally defined as big tech. If you looked at the number of employees that were added every single year, it was growing every single year.
So new employees, the number added was growing, not just the number hired. So if someone was hired and someone was fired and added, that was just added to zero. And then on top it was growing every single year. Then it dropped when it was the large layoffs, whenever it was a few years ago. Now for the last three years it's been almost zero. So for the last three years in all of big tech, it's been the number of people in are in, there's no net new people, meaning if someone leaves they'll replace them, but it's basically just there.

Yeah, I think, go ahead. I mean the famous or the example that reached the headlines was Klarna. The Klarna CEO was very,

They rolled back on that though. Did you see that

They rolled back? I did not. They rolled back on that. Yeah, they're like, oh, AI's not there yet. Oh, okay. Yeah, no, but just for context. So they announced this that they were not going to add more people, they're just going to naturally see it, natural attrition, basically try to do everything with ai, let people stay, start measuring revenue per head as the key benchmark and then let people just quit when they want to quit, but then it would not rehire for those positions.

So that part I wasn't talking about, so sorry, I'm glad you finished that. The part they rolled back on, they said there's like a fire, all their support, it's going to be all ai. They rolled back on that. It's like, oh, we can do a lot with AI internally. We use AI for support as well, but we also have humans as well because it's not there. So it does augment people, it makes them more productive, but it's not a full on replacement for sure.

Yeah, no. So I was referring to their culture that they're changing this AI first culture and we have, there's all these CEO memos. You have the Shopify memo that you need to, if you want to bring on a new employee, you need to motivate why this job could not be done with AI first. And I think there was a Microsoft layoffs a couple of weeks ago and where they attributed it that 30% of the code base at Microsoft now is written by ai. So I mean it is definitely moving in that direction, which comes to the darker, I guess, part of this conversation. It's like, I mean when you have agents, I think a lot of people's like what type of jobs are going to be augmented versus replaced? Do you have any thoughts on those?

So sure, yeah, I was just going to finish up on that where Oh yeah, yeah, go for it. The coding one. No, no, I'm going to get to that. I don't want to talk about the talent carpet anymore. Whereas coding is one where it's very useful. And so you can see that basically if flatlining in the amount of people that are being hired there, not that there's not more jobs, they're generally jobs we're hiring, other people are hiring as well. But on net there's less that exists versus the tire company. So it's like where is ai? And AI is really good at coding, and that happened to be, it wasn't on purpose, but it found out it's actually really, really good at that. So that's where it started. The bad ones, again, the ones that are most effective are generally support. That is the number one.

It seems people argument argue that sales will be, I'm actually a believer that a lot of the sales job will be augmented, but actually because of, if everyone has the power of having an AI SDR or a sales representative, then the value of them reaching out goes to zero. So what that actually means is if every company is a competitor to Infinum has a infinite AI spamming me to sell me, I'm not going to answer any of those calls or any of those emails because there's just so many that there's no value. So then in that scenario, I think an in-person sales actually makes sense because I will actually now reply if you sit in a car on a plane and come to me. And so I think a lot of the things that happened in society historically will happen with AI as well. Where it's like on one side where you're either going to be super artisanal, might not be the worst for sales or super high volume, it's the same thing with food. You basically, you'll go to McDonald's or go to a Michelin star restaurant.

Yeah, I like to say that you're either a wonder Bread factory or you're an artisanal baker.

Exactly. And I think artisanal baker still exists even with AI and AI robots. I believe that there is a market for artisanal bakers because you'll still want this one actually a human made in that sense. You want that. And the same thing with sales like oh, so you need where there's that feeling or there's a lot of empathy or you actually want a human to interact with. A lot of people go to a restaurant because of the waiter, because of the bartender, because of not the specific person, but because there is a person to interact with and you want to interact with a person. And so those things are, I believe in the long run will still exist. I'm not saying that there won't be a AI that'll do that. Surely will. And people use, I think it was, what was the numbers? I think like a human companion is the number one use case for ChatGPT right now.
And so that could be also a psychologist, a psychiatrist, and so you can talk to them, but I would bet that if you had the time and the means that you'd still go to human, you would still go to a human to talk like psychotherapist versus that. And so there would certainly be a market for that as well. Basically where you want a human interaction is something that will become even more valuable if you are in the top in what you do. The other thing is, as we talked about earlier, even in coding right now, you'll delegate the simple tasks and even do something more complex, but the actual thinking, the large view of the architecture of what you're trying to build, and also people don't think about this very often. Agents really don't do anything net new, net new. They don't do creative coding, whatever. So the product that we're building right now, so we're building infrastructure for agents, a lot of the things we've built ourselves agents can help with, oh, now that I've broken down to this little task, it can help me do that. It can help me fix the code and whatever. But if you're going to send it out to go do the whole thing on its own, it actually can't do that. What it can do is a copy of Airbnb because Airbnb already

Exists.

And so if you look at all the designs and the templates, you'll not see a new creative. You're a creative, you can see just by your style. And so it's very hard for me to believe or what I've been seeing that you'll see net new art, it can be decomposed, but you can actually see a theme that they're doing and it's not a new thing. And those things are still human as far as I know. So the agents basically, and I don't know when they surpass that, but for now it's like what have humans done? We can replicate that to various degrees, but net new is something that they don't do. So can you do something on the forefront of creativity, whatever that might be, or something that needs empathy or human or very, very long deep thinking to be able to create something is I'm being ambiguous, but I mean those are the traits that AI is not good at.

No, I mean I think I agree with everything you just said. I just got to, we need to move on for the next session. That production team let me know that there's a lot of viewer questions.

Okay, cool.

There's a lot of people online that want to ask you questions. So if you don't mind, let's go ahead. So here's a question from Alex who asks, can you run agents in parallel to split a large problem down into smaller, more defined problems, potentially having them integrate with one another to solve collaboratively? That's a

Good, so that's something that people are working on and might work now, might not. So we have even a protocol that was came out recently was called ATA, agent to agent, where you can basically have one agent, they can talk to other agents to do that. So in essence, the idea is yes, that is the idea, even though it might not be perfect at the moment where you can sort of define something and then the agent will break it down into multiple and do that. I don't think that subjectively, I don't think that works really well yet in practice, but that is definitely where we are trying to get to.

To add to that, I've been trying, I mean we can leave that in the notes, but I've been seeing some tools now where you have two or three different roles. So you have a product manager agent or you have a creative copywriter agents and a designer agent and then they instruct each other to design websites or marketing materials,those things. But I believe the question was more in the sense of the context. So because it can't do something very large, can it break it down into chunks and do that? And so I think to have that at production level, that works really well. I think we're not there. I think you as a human still have to do all those things and get that. But yeah, that's definitely where we're going and I wouldn't be shocked if we see that very soon be a way of work.

Thanks for that one. Next question comes from Pavo. What do you think is the most common misconception about AI agents that businesses should watch out for

Common misconception? So the most common misconception might end up being true, but in the sense of, I think people also very aggressively think the ones that believe in AI agents, that it will completely replace people. I think the quality of the work they still do is really lacking and it's not there. So I strongly believe in a world where it is definitely augmenting, so you as a human will do that, but completely replacing even a section of work. So if you give it a full end-to-end task, my criteria of quality, I've never seen it live up to that. So no matter what it was. So I don't think that we're in a place for that. I think that's the biggest misconception that it's going to do. That essentially can an agent go out and solve a whole problem? And if people are sort of tooting the horn that is there or that it's about to be here, I think that's a bit more out than people actually perceive.

So follow up question. So when you say far out, is that five years, 10 years? I dunno,

I don't know. So it's really hard to say just because, and when we talked about this earlier, it's like how fast are these things evolving? So there's two things. There's two ways that these agents become more useful. And so one is the models itself and the tools that the mallers have. And I also add things like deep research and whatnot as tools because the fundamental model, the LM hasn't changed, but it's how it can work with the world. And I think those two things, as they progress, we get more and more out of it. And so how fast do both those two things progress? That's how fast these things change and it's very hard to tell how fast they would progress.

Gotcha. Next question from Cindy. What are some underrated use cases for AI agents? Underrated use

Cases? I have to think about that. What are underrated use cases? Underrated use cases. So I dunno if it's underrated, but I haven't used it till quite recently. It is to interpret the text that I'm going to send out. So I'm sending out an email and then I give the agent, if you were X person, Y person, this I give as much context, how would you interpret this email? What emotional effects? Does it come off strong? Does it come off demanding? Does it come off soft? What do you understand from this? And so using that to understand what the other person will read from that is something that I only recently have done. And I do that with all important common communication right now. It's like what does a person understand? And I also, when a person sends me an email, verify that what I understood, the agent also understood. It's like, hmm, did I understand this correctly? And then I said to the agent, what did you understand? And then we try to do that.

Yeah, yeah, no, it's my favorite way to use ChatGPT. I have virtual assistants that model different authors or stuff and then I have them kind of proofread what I'm receiving or what I'm about to say. Like, Hey, am I coming up passive aggressive here?

Yeah, I don't want to do that. It always says yes, so then I need to change it. Yeah, exactly. Okay, so that's actually everything that we've gotten to the hour. I was thinking in the beginning it's like, oh yeah, this is going to be a long time, but we are already at the top of the hour actually, so we need to wrap up. So final question, which is kind of like a staple question of this podcast, which is called Delivered. So in wrapping up, so what's the one thing that you have delivered in your career that you're most proud of?

So I was going to actually do a trick one on that. I saw that the tagline of Infinum was never stop. And so I don't think I've delivered the best of my career or life in all honesty. So I've done great things like I created a company, sold a company, created another one, had a child, which is probably one thing that you can probably, it's really weird when you

Have, I think I heard something in the background.

Yeah, you did. Yeah, probably. Yeah. So a year and a half year old, which is great. And having another as well. And when people say, oh, the kid's the best thing I've ever done, I can understand when they say that now. But as far as career arc, everything that I've done and my partners have done throughout their career, everything's built up to this point. And I think that we are on the verge of bringing something actually quite important to the world. I know a lot of people say this, but fundamentally we plan to be the home of run everything that agents actually do in the future, which is quite large. And so I think that is the thing that we're hoping to deliver. And so dunno if it's acceptable answer, but that's the one I came up with.

Sure is. So for people who want to know more about you, where can they find you?

So to find me LinkedIn, Twitter or X if they call it right now. So Ivan Brazen, first name, last name also like daytona.io is the website. You can check out our GitHub. Anyone that wants to use Daytona as well as open source, you can do that. We have a Slack channel for the communities and I'm also in there, so you can DM me in there. So basically on the socials you can find me in our Slack and the website. GitHub. Yeah.

Very cool. Well this has been great, Ivan. It feels like we could have had another hour or more. It feels like we just scratched the surface.

We didn't get into the dark stuff, so that's good.

Well maybe we'll have you back on to talk about that.

Yeah, sounds

Good. Well, so thank you for joining Delivered. Hope you have a good time in Split and a nice trip back to San Francisco.

Thanks so much. Enjoy the conversation. Thanks for having me.