РАСШИФРОВКА 🔗
Tomas: One engineer, one mission, one month. And so this is how we are how we've been operating for many months. We are adapting, we are we are changing it a little bit, but this is the new model. And I think it's changed a lot.
Sebastian: Welcome to another episode of our new season of Beyond Vibecoding, partnering with Impala Search.
André: The go to tech and executive search agency in Germany.
Sebastian: In this podcast, we explore the transformational change in software engineering and knowledge work in general. I'm Sebastian Heidemeitzer Appen, CTO at Ecosia.
André: And I'm Andre, CTPO at Trusted Shops. Great to have you back. Spoiler alert, if you want to understand how far you can go with agentic engineering, this episode is for you.
Sebastian: Exactly. Our guest today is none other than Thomas, CTO of Omio, and actually he happens to be my former boss because I was also at Omio in the past. And he describes the fascinating approach to genetic engineering at Omio. just to drop a few facts, month they have monthly missions, one engineer, one mission at a 100 plus strong engineering orc. Insane amount of output. And since the episode is so packed and we had too little time, we already agreed on a follow-up where you can actually send us questions for Tomas. So listen carefully and send us your questions afterwards via LinkedIn or email. You can see the show notes. It is absolutely worth listening. Welcome to our episode, Thomas. we know each other from the past. that's why great to see you again after I don't know, we we talked about right seven or eight years or something. So super cool to hear what you've been up to. And yeah, please to those of us who don't know you or to our listeners, introduce yourself briefly.
Tomas: Yeah, thank you, Sebastian. Thank you, Andre. Thank you for inviting me. I was really looking forward to our discussion conversation today, especially because we knew each other. So it's great to be able to meet again in this new environment. And yeah, about myself. So I'm I'm probably a software developer. I've been one for many, many years by now. I've been developing software in large companies, I've been developing software in startups. I I spent my time at Microsoft and during the last eight years, actually nine years, I've been a CTO of Omio in Berlin, where we have been building successfully, I would say, a global distribution platform for ticketing, plus number of other engagement and functions, but this is the core all we do. So we build a platform that connects transport providers, be it airlines, train companies, bus companies, ferries, to one distribution system where we normalize and then distribute, allow people, our customer, to search and book tickets. Anywhere from US through all of Europe to Southeast Asia, Japan, Australia currently, and we are going further and further. So currently I think it's fifty-six countries and growing. yeah, and we also have a B2B angle to that. So we're also working directly
Load full transcript
with companies where we also allow them to add a ticketing distribution of transport options to their offerings. And we're working with decent companies, if I would name Uber. We are working directly with Google, we are working with with many others. So it's quite a interesting, quite complex product and I really enjoy that. Yeah. And been building it for eight years and going on it.
Sebastian: Yes, that and it's an amazing journey. So having been part of Omio's journey just for a little bit, right? I also really appreciated the complexity of the model. in the like the integrations are very like different, right? And then the the life cycles of the different kinds of data are very different. And then yeah, so it's it's a very interesting product and yeah, great to hear that it's moving in the right direction. Yeah, cool. Thank you for the brief introduction and like usually we also wanna ask you how you utilize agents in your day-to-day work.
Tomas: yeah, heavily, heavily. So I would say most of my work nowadays happens in Codex. We are open AI partners, so it's more Codex than Cloud, but I don't think there's much difference. And pretty much most of what I do is with one of the many tens, maybe close to a hundred different like Codex projects, be it development analysis. responses, updates, thinking things through, responding to stuff, etc. So Codex is the the main tool that I'm using and there is a number of agents in there. And on the personal side, I also built a number of tools myself recently. I super enjoyed it because and it was it's the change that happened about a year ago pretty much I guess for most of us I'm reading also about many people who are After years and years getting back to personal productivity and starting building tools.
André: Yeah.
Tomas: And I I I told you, I told you Sebastian, one of the things I'm working, let's say Omio has like currently four major sites. It's Berlin, it's Prague, it's actually Bangalore that we launch in Singapore. And I'm mostly traveling I'm having lots of ideas. I have a time to think about, and I built a lot of tools where I can interact with my agents I talk to them, they respond back to me. I can process many things, many information, I can take notes,
Sebastian: and and how do you do the the TTS and vice versa?
Tomas: So I'm using the OpenAI real-time voice processing currently. These
Sebastian: Yeah, okay. Mm-hmm.
Tomas: are perfect, honestly perfect. I used different things before, but probably like months ago, one and a half months ago, when OpenAI rolled out the latest real time processing. That's that's top notch. It's super fast, quality is great, and I love the ability to switch languages and go from from English to other languages and back. It's quite cool.
André: And you decided to build it on your own because of like to get your hands dirty or because you you're missing that specific functionality.
Tomas: W I'm usually solving particular problem.
André: OK.
Tomas: So and this was really like I was I felt like something like this needs to work and I was initially I played with Claude and I played with OpenAI and I found it not not good enough for the particular use case. So I then said, Okay, how hard it is? And then and it's not that hard, honestly. It's just I have my word, but You can run it, it's it's it's so fast, it's so so powerful, so I really love to do it.
André: This is actually a good question. Back in the days, was this quote or there was this quote, this can be so difficult, right? I'm wondering if that became true now with AI because it actually is simple now, right? Couldn't be that hard. It isn't that hard anymore. Yeah. Yeah.
Tomas: I am fully converted. I can say that. It's not that hard anymore, period. I think complexity, of course, there is complexity, but it moves to a very different level.
André: Yeah, yeah.
Tomas: But I think this is how hard it could be, it's it's generally that's that's the wow effect, maybe where if people try really. and spend minutes on it, not hours or days. They usually they are like super surprised how good the setups are today and how powerful the tools are.
Sebastian: I think that's true. Yeah, I see the same. Yeah. Awesome. That's really interesting, really cool stuff. And yeah, I have to admit that I also have my suite of of tools that I built myself. It's not that much that I have my own interface actually. Most of it is actually going through OpenClaw or other interfaces.
Tomas: Mm-hmm.
Sebastian: but getting there. Getting I'm currently building my own interface actually for a very specific purpose. All right. cool. Then I think it's time to move on to our main section, right? Which is as we call the mead. And what else could it be if we talk to the CTO of Omio? Then the engineering approach, the the agentic engineering approach at Omeo. Yeah, maybe you can yeah, talk us through a little bit how you do it at Omio and yeah, then we engage in a discussion to go deeper.
Tomas: So I think, Omio we have a lucky setup, I would say, where AI and becoming AI native is a company priority and has been a priority. And we've been really working on it since last year and definitely this year, and we are going through this transformation constantly on an ongoing basis. And I see it fascinating and I see it very very interesting and with great outcomes because as the tools are improving, everything keeps changing and it ke gives us more and more power and enables us to do more and better on an ongoing basis. And it's amazing to watch things coming up left left and right. Honestly, there are many tools which help help us work better. So we apply AI, I I always think like on three areas. One area is obviously just craft, yeah, software development. Like how do we build software? How do we write code? And how how that happens and how that happens faster in better quality? And yeah, that's pretty much it. We need to be faster and the quality needs to get better. Then it's about
like can we change what we do internally by using AI tools and are we building tools to actually enable much better operation? And definitely this is something we are thinking, we are investing, we are doing a lot in that department. And then we are looking at the products, how can we actually help our customers with these tools and what what can we build so that our customers can benefit from that. But I think the main question is more about internally how does AI change our work in OMIO? So I think we have a number of I would start with internal product projects. Yeah. Maybe Sebastian will remember, there are some quite complex exercises in our business or task, and that's integration. And for us, this is we are integrating with many, many, many distinct systems. Every partner that we work with is different. You can say there's no two external partners that would have a similar interface. And what we mostly do is we are building interfaces that translate complex models on the side of our partners into complex models, which is on our end. And I think we managed to build automation or actually agentic system which looks at that autonomously and is able to build most of most of these very complex integrations in very short amount of time. And work with through feedback, clues with engineers, etc. And this gave us enormous boost in how quickly can we move forward, how quickly can we further integrate and expand our coverage worldwide. And that was, I think, one of the major achievements. And then from the other parties, the system is for the same reason our systems are very complex. Probably Sebastian will remember. It got much worse over time, you can imagine. And
Sebastian: Yes.
Tomas: we actually invested and built a number of internal tools which are absolutely capable to look at the bug report, analyze the analyze the problem, analyze log files, look into various things, kibanas, grafanas, whatnot, reason about this, propose a solution, propose the code and goals. nearly up to the PR which then needs to be reviewed. So our kind of ability to resolve problems by applying these tools gets into next layer, next level. So we're getting getting to the point where yeah like engineers are just managing the outcome of agents or farms of agents and just validating and ensuring that whatever comes back is is good enough and then releasing into production, of course, test is being automated, test are being AI automated across front end, back ends, etc. So pretty much all the process, we are introducing various types of agents at every every stage. And it's it's really changing how we work and it's making it I think better and more manageable and fa and faster. So that that's what we experience. Then engineers, engineering work itself, of course, everyone has codecs. So we we normalize on codecs. I know there are some people who are heavy, heavy still clutching to clot code, not saying anything, is the same tool anyway. But mostly we are we are running on
codecs. And pretty much all the engineering, all the code production or building stuff, building services is moving into codecs nowadays. And we can talk about later where where I think it will end up and where is it going. But that's it. So we are allow we are enabling people with tools. We are constantly also working and making sure that we are efficient in their their changes. But it's a it's a fun journey, I would say. Very fun journey because the tools are improving on monthly basis. every month something better comes. And it's it's great. At the same time it means we are still learning. We we have to introduce new things. We always look at is it this model, is it that model? Is it dead harness that harness? What's what's currently top top in the world? And we are trying to bring the top tools and top kind of support into the hands of our engineers. And what is Also amazing that it's also changing the work of product people and works of how design works and product works. And what I see is a little bit of blend. Yeah. We we got from mostly from writing down product specs into building prototypes and implementing prototypes mostly internally. Our designers are coming back with actually prototypes which are working the way they expect to. They don't need to do wireframes, they build apps and the apps are are being then implemented and taken into another spin. But we we have very, very good and fast way how to bring the ideas in and bring everyone together. So we don't need to translate or write like complex complex product definition files or anything. So everything is becoming way more understandable
André: then.
Tomas: and I would call it fun. So all these parts of of the whole engineering process is improving and accelerating. Which also affects
André: No, I have-
Tomas: sorry, I can talk I can speak for another hour, but stop
André: Yeah, maybe we take a break here and then maybe try
Sebastian: Yeah.
André: to dig deeper on one or the other topic. I was wondering, Sebastian knows this probably, but maybe for me and the listeners, how many engineers, how many tech people are there at Omeo just to get a rough understanding how complicated it also is to align everyone? And that would be my second question on one workflow or how flexible do you keep this for the tech people?
Tomas: Yeah, it's a good question. Again, I can go off another hour on this. But the
André: Hahaha
Tomas: the size, I think we are pretty much similar in size to what it was this eight years ago. So it's about 150 plus minus engineers
André: OK.
Tomas: and additional product and design people. So currently I think we are running like what we call TPD, around 200 hertz. And that's
André: Okay.
Tomas: pretty much where we sit. On the process.
André: And are you... Yeah. Yes.
Tomas: I think we we s we started and we always were agile, I would say, Scrum company. Scrum was something which I really liked. And we've been running bi weekly sprints since ever and quarterly planning with with six prints in there, and that was our mantra. And we changed that this year because we found that this is not really sufficient anymore. It's too slow. And And maybe again, Sebastian, we we went through actually if you think about our planning process. So we used to do bimonthly planning in the past. And it was very, very hard at the time because just to create a plan took more nearly more time than to build a problem, etc. So then we went to quarterly planning, which worked much better because we actually we could product could spend time thinking about the next quarter, and then once it was defined. The engineers were able to deliver and somehow it was coming together nicely. But this was how we worked until beginning of this year, but then we decided this is not this is not working in this AI native or more AI AI enabled world. And what we are do now, what we call is monthly mission planning. So we run in monthlies because we saw that our engineers are capable to deliver in one month. themselves with with all the agents and all the AI-enabled environment they have, they deliver pretty much the same what what the small team used to do a quarter. And that's something that we also needed to somehow bring into our planning. So we are doing so-called monthly missions, and mission is whatever you think about, it makes it's it's a project, it's a idea that needs to be Interest sufficiently interesting to be built and needs to land in months. And most of what we think about currently is doable within one month's time frame. And so and the idea has one what we call DRI, directly responsible individual. It's one engineer, one mission, one month. This is our model currently. It's totally cool. One engineer, one mission, one month. And so this is how we are how we've been operating for many months. We are adapting, we are we are changing it a little bit, but this is the new model. And I think it's changed a lot. First, we much better understand what we're building. We have certain kind of common expectation that actually works across all teams. And this and that was a question whether it's Yes, we have teams, we have backend teams, we have front end teams, we have acquisition side of the house, we have CRM, we have postbooking. Actually we see that this kind of mental model works pr pretty much across the board. Anything nowadays you want to build probably will take about a month. And I don't mean months of coding. It's like first probably there is a mini hackathon at the beginning where you try to understand quickly, get the prototype in, get the feedback. Then you probably next two weeks it gets to the position where we can actually bring it out to customer or user or whatever that is. It could be service, it could be front-end components, could be new feature. Then we iterate on it a little bit, but everything happens within one month. And in most of the cases, and again, world is not black and white, so I would say most of the
cases. It gets delivered happens within the months and we can see we can see the impact pretty much very quickly.
André: I need to have, I have a follow-up question. If you say one engineer, one initiative, one month, there's still a team behind that, or are you then running 150 initiatives a month?
Tomas: We are there are teams behind that, but it's much looser kind of concept than it used to be. It's really it's really delivered by engineer. The whole expectation moves down to personal individual delivery. And that that's a big change honestly. We don't have
André: Absolutely.
Tomas: or we are like I I am frankly opposing the ideas let's have these three three people work on it, because I don't think we need to. One one person with right set of agents with the tooling with the harness, they can do it. And it's more about a little bit top-down expectation setting because it's scary. It was scary in the beginning, but in the end it's learning exercise. You do it one month, you do it second month, third,
André: Yeah
Tomas: fourth month, it's pretty much works. So teams are there more about
André: I'm,
Tomas: the concept of team for me is more more a unit of ownership. Because you're still building things, you need to take care of things, you need to keep things working. So teams generally work certain areas of the system. They understand how it operates and runs, but the engineers, part of the team, they can even float around the organization quite well right now. We are able to actually doesn't mean that one team has to always do one thing. They can get help from others. We can group people to deliver bigger problems by having multiple missions, or the people from one team help others. And the whole work becomes fluent. And I'm saying it because my past experience was very much against this. Because in the past, you would not want to have engineers like running around because the onboarding. Time to learn new. Concepts, even new languages, you run multiple languages. Well, we were always limited by I would say technology and by deep understanding. And I think this is being now handled by the AI agents, honestly. So for me, computer
André: hm?
Tomas: language is not the argument anymore. Anyone who's coding, whether they are going doing Golang, they are doing JavaScript, whether they do I don't know what.js, JavaScript, Java. same thing people can onboard th onboard in days nowadays so it
André: Yeah, I would agree. was maybe, so in this regard, totally agree. Nevertheless, I would say if you dynamify an environment, you actually need to maybe provide other conditions. don't Good architecture documentation, high test coverage, these kinds of things. So absolutely agree, like programming language probably is not an issue anymore. But to be able to deliver, could imagine, but curious to hear what kind of changes you took to, like in your environment, whatever it is, to make that work, because sounds pretty fascinating to me.
Tomas: Yeah, it makes sense. Even so I'm I'm my mind or my opinion this is a little bit changing, but makes sense. We started by building claude MD files to every report. We started that way. Then we look at OKF, open knowledge format type of information, building and collecting the information that people but mostly the agents can utilize very well to understand the areas. And then I think we've been building and we are still building the the the overall knowledge base of everything that happens in Omio. But agents
André: Hmm.
Tomas: themselves are very good to actually read and learn from various sources. So it's more about providing right pointers and documentation of not on top of that.
Sebastian: I I will have tons of questions. I fear we probably need a follow-up because there's so much to unpack. But maybe and we we don't have that much time left, but maybe two two two questions. So number one is more on the front end of things. So how do you integrate with PMs and designers, right? Because you mentioned the DRI is an engineer, so how do they collaborate with PMs and designers, right? And then more towards the back end in the way you describe it, right? You have a massive increase in productivity, right? Because what was done by a team in three months before is now done by a DRI and you have 150 n engineers, yeah. So there's a lot. So how how has the the productivity pretty much increased in in the sense that how do you see, I don't know, the the amount of PRs being merged or are there other metrics that that you can see that that yeah pretty much show how how much the productivity increased.
Tomas: So I I don't have these numbers handy because I don't look at these that much, but the increase definitely there is a three, five X, ten X increase in these over the past year. That's for sure. I'm more and we are I'm trying to make sure that we are looking on the real outputs. So I don't care how many TRs it takes, but what I care about what does it generate? does it really work for the customer? so I'm more We are more interested in product APIs, product metrics, which at on at our size it's it's the best answer. If the conversion improves, people like it. If the conversion tanks, people don't like it, and we need to work harder on everything. So we have a set of kind of product APIs on all the stages and across the stages of the product that we are very, very religiously looking into and evaluating. But going back into how the product and design work. So product, I don't think much changed. It just made them more important. And way more important because yes, I I still see a this distinction between product and engineer, of course. Like the product people, we need them to actually think ahead. Yeah, think about what What would make sense? What would make what will work for our customers? Of course, what is happening in the competition? Where are the opportunities? Right? Really? We need this what and which is mission specs. And that and we need more of them
because as you said, by the way, we have 150 engineers, but you you have to keep the lights on. So probably we
Sebastian: Yeah, yeah.
Tomas: are running about hundred, hundred ten missions monthly. It's more like the hundred twenty is the number. But
André: This is massive.
Tomas: it is. If you see it. Yeah.
André: Ok, sorry, don't wanna interrupt you, but this is still massive.
Tomas: And you need to you need to get the missions back. So but we don't need, I don't know, probes. We don't need 100 pages. What we need is really some kind of, yeah, that's that's what we need to achieve. This is how do we want to measure our success. These are the key KPIs, these are the outcomes that we're looking after. And that goes to the to the mission and that drives the mission. And then the engineers they have a they have a run on it. Yeah. So but I would say definitely products is more and more important because as the velocity of engineering proves, you need to feed it. You need to and fortunately for us, there is we still have a lot endless amount of problems in our backlog that we that we
Sebastian: Yeah.
Tomas: want to go through. But yeah, it's definitely so product, what I what we what's improved, it gives I hope it I think it gives them a chance to look more forward. And then don't need to go too deep into definitions because that's more understood and and works better. So that's improves their life as well and of course they use codecs to do analysis, to look at data, to process data. We have enormous data warehouse with numbers that you can imagine. And again, we build tools which allow everyone to actually query it in some various way in quite intelligent and reasonable way without needing needing a diploma from the area, while we have people who can go that deep and they are still absolutely needed. But most of the questions we have and most of the questions we want to get data is not that complicated. And it's enabled by our other data agent, which works everywhere. And I use it every day. Everyone can use it. So it's very cool. So that's that's on the product. Design, I think design is again super, super important. And one thing which is still missing and which will which we need and will be needed from humans, and I just wonder how long, but it's it's a taste. Yeah, because the agents they don't have a good taste for for how how the UX should look like, how the pr how the things should be presented. And this is super important and super differentiating for the companies. And I think that the design plays totally critical role because they need to drive the look and feel. Look and feel, or even the whole flows, but that's less of a problem. But the look and feel and build a taste and build a brand and build that. So that's I don't think anything is changing, but again, their life is easier because they can actually they can build prototypes. And they do. They build it. They use the tools. Now they are moving into codecs. They started with lovables and these things, but after some time you go into and I just build it in codecs. I just build the Android app. And I build another Android app and yet another Android app. And just iterate on it so that it looks good. Give it to engineering. I want this in the product here. This is how it looks like. Go build. And you give the app to the agent, and the agent looks at that and goes. Simplification, but I think that's that's what it becomes. So it s simplifies
Sebastian: So it's basically basically yeah, basically ahead of the month the mission specs need to be prepared, right? So that the DRIs can take care. And then there's probably also some verification or so at the back end
Tomas: Yes, yes.
Sebastian: that involves BMs, right? But then in between it's largely pretty much DRI. Yeah. Yeah. Yes. Yes.
Tomas: It's that is a conversation of course, PMs to engineers and reviews of course. But but in a nutshell this is. And what is really amazing that we are every month, we just finish one month and then regroup and it could go into many directions. And we are able what I like, we are really able to react to the needs and things that are happening in the market and things are happening at again. at a unprecedented speed and this gives us and I think pretty much anyone to actually ability to really react and be somewhere in the front of the whole story, of the whole development.
Sebastian: Amazing. Thanks for sharing, Thomas. I'm afraid we're already at the end of the episode, but I'm thinking about a follow-up. Let's let's see if we can get that in because it's so many questions. before we finish the episode, we usually have two smaller sections towards the end, which is the reality check and then the prediction, right? So on the reality check side we ask our guests always if they have what the fuck moments or wow moments. So yeah, do you have something to share with our listeners in these categories?
Tomas: So less of WTF moments. Of course, we have our share, but maybe I don't want to talk about them in public. But it's not it's not nothing new that much. It's still some things, it's more like maintenance coming back and hitting us two years later because we didn't deal with something which we should have dealt, but because we don't have time and this and that. Wow moments, these are in mostly I would say in the how quickly and how well things can can go out and be built. And maybe my biggest wow was we had a hackathon and we have a history of running hackathons and we did one early in the months. And unlike before, like this time, as two days in two days, our teams built probably close to again sixty seventy projects, maybe less than that, sorry, I'm just But we have a good thirteen things in progress right now. Things that are really going and coming and we absolutely say, Yeah, this absolutely needs to ship. And we are bringing it into into missions and we are bringing it out, and that's that was not happening before, and I really love the quality and then even wow was like how how well they can present because they use the agents to do presentation. Sometimes you feel maybe a
André: Ha
Tomas: a little bit of more human touch would be better, but at At the same time, the quality is insane, this thing that they can do. And it's being done again by very small groups, two, three guys, one within two days, and it's amazing. So I think the speed and how much how much is being enabled and how it works is probably what really surprises me again and again.
Sebastian: When you described your current approach of these monthly mission plannings, but pretty much it almost sounded like a hackathon. So it sounds like every month there is a hackathon because there's a new project. And I mean, in initially when you said how much it's automated and and whatnot, right? I was thinking, Okay, how does it feel for engineers? But engineers love working on new missions, right? So I I think it's you you're threading the needle pretty well. I I yeah. I d don't know how your engineers feel, but but it sounds very interesting. Yeah. All right. Then last but not least, the prediction. Do you have a prediction for our listeners?
Tomas: Right. I think there are everything the role of engineering, software engineering is changing. We can see it already. We we talked about multiple aspects. And interestingly, I think that it's blending with the management. And we see what we see and what I see and what we even talk about internally, but also from the outside is that the most successful, I would say, AI native engineers or internal hackers are ex managers. People who actually have experience with managing people are then very good and managing agents because they can delegate, they can deal with, let's say, lower than expected results, they can deal with different solutions that they expected, they would do themselves. And and they are really very, very much more productive and they multitask and these things, and these are all the things which are not so natural, but I see it coming and it will make a difference to to how people the best operating or best performing people will operate and and work and it's another kind of wave of change which is very interesting. But I think my take is that software development moves from writing code to managing group big groups of agents which are doing the job and it's about feedback, it's about goals, it's about Even performance reviews of the of the agents if I if I work you need to navigate, manage the group and let them evolve and that's probably what's going to happen. Flatter works, most of this kind of operation and much higher output overall in this
Sebastian: Wow, thanks a lot, Thomas. It was great to have you on
André: Thank you very much.
Sebastian: the podcast.
Tomas: You're welcome. Thank you very much as well. Thank you for having me.