What GitHub's Ex-CEO Says About the Future of Developer Skills

12,101 words · about 48 min to read · YouTube’s automatic captions

Brainstorming with coding agents before writing code

0:00 If anyone who's predicting what looks like in a year or two is going to be fundamentally wrong. We don't know where we're going other than that we do know that we're going to keep building software while maintaining billions if not trillions of lines of code that are already out there. A lot of these coding agents are actually really good at brainstorming. And so a lot of my developers are using cloud code to write code but before they do that they also use cloud code to brainstorm. And I don't see that going to be replaced anytime soon.

0:25 How do you envision the job of a future AI native developer? The biggest skill as a software developer as an engineer is that you know I think you're going to see more of a agent to agent collaboration. What's really interesting about that right now is it gives me as a human feedback on code now, but if that code was written by Devon, why do you give that feedback to me, give it to Devon and have Devon figure out that is how you would do that in a human to human developer team, right?

0:48 What can you say about your new startup and what made you sort of take the leap from from where you kind of were in GitHub? I think that's the Before we jump into this episode, I wanted to let you know that this podcast is for developers building with AI at the core. So whether that's exploring the latest tools, the workflows, or the best practices, this podcast's for you.

1:17 A really quick ask, 90% of people who are listening to this haven't yet subscribed. So, if this content has helped you build smarter, hit that subscribe button and maybe a like. All right, back to the episode.

1:29 Hello everyone. Uh, welcome back to the AI Native podcast. Uh, today we have a really exciting guest that, you know, I've known for quite a while here. Uh, we've managed to juggle schedules to, uh, to make work. And that's Thomas Donke, uh, who is, uh, who until recently was the CEO of GitHub, uh, and is now we'll talk about this in sort of the brave AI dev, uh, sort of startup land on it. Thomas, thanks for coming on to the show.

1:52 Hey guy and thanks so much for having me.

1:54 Um, so uh Thomas, it's it's sort of funny a little bit as we connected here or like know reconnected. we've been sort of chatting throughout the time. uh talked a little bit about uh how we how we met uh initially and this was was actually like it's not sort of hasn't been publicly discussed but there was a moment in which it looked like GitHub might acquire sneak uh early in the journey uh and I guess that's how we uh we got in a room uh and you know clearly clearly kind of led to uh to interesting paths as uh as GitHub got pretty big in uh in sort of apps space on it clearly didn't happen on the snick side which worked out pretty well uh but uh yeah those are like fun and intense convers conversation. There was a way to start a relationship there.

2:33 Yeah. So for the you know for the listener to have some context uh it was summer 2018 to my memory. Um so Microsoft already announced uh that we would be acquiring GitHub. Um I was part of that acquisition team together with Ned Freiedman who then became the CEO of GitHub and a few other folks. Mario Rodriguez who's now the head of product for GitHub. And as we had signed the deal we were like thinking what else would be a strategic for for the GitHub portfolio product and our platform. uh to to be added to and we had a long list of interesting startups in the space and and one of them was was Sneak. So to my memory we had you over in uh the Twitter building uh funny enough where where back then Ned's office was um and in the conference room we chat a little bit of

3:15 you know how your business is going and uh how the you know open source security space is important for GitHub as a platform and obviously for for sneak as as a product and that's how we how we met the first time.

3:25 Yeah. Yeah. I remember I like remember the conversation very vividly on it. We also we dragged Nat at some point I think it was before that actually to come speak at our all hands uh at uh at a sneak all hands as he was sort of exploring possibilities and it was interesting you were in that sort of annoying window in which you couldn't act you could just explore because you were sort of waiting for the GitHub acquisition to uh to close and in the meantime investors swooped in and uh and gave us like you know higher valuation uh opportunities in it and

3:54 and it was sort of been fun times clearly both kind of worked out pretty well like you know GitHub built a thriving security business. Sneakers have grown well. Uh

4:02 well, you know, if you could time travel back to 2018 and tell both of us what the valuations of developer tools startups are in 2025. You know, in 2018, everybody thought the 7.5 billion price tag for GitHub was high and the valuation for Sneak was was growing quite fast and there were a number of other companies that felt quote unquote overpriced. Uh in in today's world, that's that's almost a seed round.

4:25 Yeah. Yeah.

4:26 And uh and um I think you know there's so much excitement and we'll get to that later in the part but there's so much excitement in the space that um uh people now believe that developer tools companies can become much bigger than than back in 2018.

4:40 Yeah. Yeah. I fully fully agree and I guess kind of you know part of it is the sort of the software is hitting the world but also just a bunch of uh you know like more developers and and just sort of better learnings about how to turn I think a bunch of these tools that were just described in like technical terms into uh into kind of real business value right and uh so maybe it's just the software is eating the world kind of trend you know manifesting uh and it's all inevitable. So I guess you're you're back in the sort of the uh the sort of the startup uh reality of it you know

Incumbents and startups operate under different financial pressures

5:10 after uh after GitHub um and you know a lot of the conversation when I sort of hearken back to to that moment was on kind of startups versus incumbents right you sort of have the large companies kind of coming along have all the distributions you know in the security context for instance GitHub eventually acquired and launched dependabot uh on it and clearly you can now kind of you know flip a switch and suddenly all the repos get uh uh you know get this functionality which is this like massive power on it. uh and and yet kind of kind of common wisdom is that sort of startups are uh are where innovation happens, right? And you kind of you learn the breakthroughs happen in sort of startup land on it. The incumbents come along and I was curious like you know you're you were sort of in the sort of the heart of it, you know, from uh from the sort of the incumbent land in GitHub. Uh and now you're you're kind of going back to startup. I guess what's what's the the view you've you've you've built up in terms of where innovation will come from in this world? like you're you is it still sort of the startups will uh will innovate or are

6:11 incumbents kind of clued in enough that they will that you actually expect uh a bunch of the a bunch of the innovations and stuff to come from there?

6:18 We see innovation everywhere, right? Like it's it's both the startup world and the incumbents kind of trying to out innovate uh each other. Um, I'd love, you know, to think about competition in business as very similar to competition in sports. Yeah.

6:32 And you have, you know, a leader, um, you know, right now, uh, uh, in Formula 1, it's now McLaren, but for like the last four years it was Red Bull, right? Right. And so that's effectively the incumbent and they have the challenge of staying on top of the game and they get the best people, you know, poached off to other teams because they have a career path there and can become, you know, the team boss while at at Red Bull they were kind of like stuck under the stuck quote unquote, you know, under the under the previous team boss. And the same is true in in our industry. There's the the big ones uh that make a lot of

1 Highlight · Incumbents and startups fight in different weight classes

7:01 revenue. And I think it's it's it's easy to forget, you know, when you're in the startup land or when you're looking at these two worlds that the the big companies, they fight a very different uh uh battle, you know, if if if you, you know, uh want to grow a Microsoft at double digit percentages every year, that's like, you know, double digit billions of dollars um that that have to be added that net new, right? Like on top of the renewals um as everything, you know, these days is a SAS business or like a recurring revenue business.

7:31 And so it's a very different game if you have to find 30 40 billion in net new revenue uh versus your small startup that is happy when they get to the first you know one two or or or 10 million in revenue run rate. I think that's number one is like you got to realize that these they're not fighting in the same you know weight class in boxing they're fighting in very different classes and the so but then you know when you look into the what they're actually building the products and the features it's it's actually in the same space and I think that's really what makes this so

7:58 interesting and you see the big companies you know all adding AI features and building agents and and pushing really hard and and staying you know in the news and then in the middle kind of like are the AI studios like like open and tropic and then there even there that's big different sizes of of these studios, right? From really big like open air to to quite small and and somewhere in the middle is Mistral and and Black Forest Labs and what have you. And then there's the, you know, companies that that are completely new and they're building with these AI tools what you might call AI native, you know, from the start. And I think that's a very, you know, uh, exciting space because effectively, you know, as as a dev, uh, you start your morning not by going into Jira or if you're, if you're more lucky, get up issues and if you're even more lucky, maybe it's linear. Uh, you you go and open a terminal and you open cloud code or or AMP or Gemini CLI and um you you kind of like brainstorm with the agent as the first step, right?

8:53 And so that's a very different way of working than if you're in an existing codebase in a large company uh where you have a lot of sneak findings that you have to work through. You have your product manager backlog. You know you have your technical debt. You have your on call rotation and so I think both of these sites will innovate in their own ways and it will be really interesting of who can you know change the world faster. Yeah, I think it's I really like the sort of the lens on the sort of the the dollar count and I think it's one of those that gets a little bit challenged I think with AI with just the level of attention, right? because it was uh our CTO at Akami uh before founding Sneak and you know as we were sort of building business cases over there it was immediately like yeah can this be you know a fairly instant within the next couple years at least like$und00 million AR business because otherwise you know it's kind of hard to really rationalize it within the the system and so that definitely opens that window for companies to to build up and prove out that a business will become you know that that size but it seems like in AI there's kind of this uh uh sort of sense that, you know, things can become a

9:55 billion-dollar business in such a short window of time that that bets are legit, you know, to make even within the incumbents. No, like, you know, I think that feels a little bit different now than it did in the in the traditional kind of a startup versus incumbent reality.

10:12 I think it is and it isn't. um you know and the revenue the revenue numbers or you know the enterprise sales motion is also often used as an excuse

10:20 you know that you don't focus on the product anymore and you're accepting that the user experience sucks um

10:26 uh you know I was I won't call out the the product I was trying to set up SSO for you know a a tool that we want to use within with a new my new startup and

10:36 uh I couldn't figure out how to do that and it's it's like you would think that's an easy thing these days to just connect your your Google identity and and have SSO for everybody. Um, and I'm sure you know somebody has a reason why it is the way it is. And the challenge that we all fighting is that we have not enough uh time in our days. And u so everything you know whether it's Netflix and YouTube, whether it's work, um whether it's you know real life things like Lego has become a a battle about uh your wake time uh and you can only reduce your sleep time so much until until things that's really unhealthy, right? battling time and at work that means priorities setting priorities and then I think again this is much easier at a startup even more so if you don't have a customer yet right like I was uh I for the my new company is fully remote so everybody is somewhere around the world and we recently met uh for the first time together in person in Singapore and one of my open remarks was like this is the going to be the most happiest time in our in this company life because we have no customers we have no production system nobody is on call you We we're just prototyping

11:39 and being creative and so looking back you know in a few years this will look like the happy time because in in a year from now hopefully you know we have you know a product out there and we have found some customers and maybe they or even paying us some money and then the life is very different because then you have to prioritize between the thing you want to do and the thing that actually is driving the business and that's what really big companies uh or companies that have found you know some form of product market fit are fighting every single day and

12:06 and Um AI is shifting that I think to some degree because agents allow you to take over things that you don't want to do you know so you have more time for the things that you do want to do. I think you know security findings you know you mentioned dependabot uh uh those kind of things will get even easier because an agent can just solve that for you to the degree where you only have to approve at the end that it's ready to deploy um you know cleaning up you know tech debt. Um, I have the the VIPE coding book here from from Gene Kim and Steve Jagger and and it has a lot of these anecdotes in there like how they're going into like I think it's 25 30 year old codebase and all of a sudden they're burning through the issues that they never had any motivation to pick up and and that ultimately then means you have more time uh to to build to build something cool.

12:48 Yeah, I think I I relate to the sort of the speed uh kind of more maybe than the business model specifically in AI like I think I think that is a change uh and and to be able to build a business in a native mode. I I love the I referred to the first sort of uh you know months maybe I guess of Tesla's run as like the honeymoon period because you know like because reality hasn't hit you in the face yet. You know you're you're uh you know you get to build those and you know but nobody has sort of told you your product sucks yet. You know nobody told you this is technically in feasible on it. I guess there's an there's an exception to that. I think both you and I were in a kind of privileged place in terms of ability to fund raise. So there's like you know not all founders are kind of in that reality. But once you pass the fund raise uh then uh then those are good. Uh although you do there's something amazing about users telling you your product is amazing. Uh and you know that's are very satisfying and so the you know not h not not having the user to tell you it sucks. It's it's you know it's like two sides of the same coin right. Well, you can't stay in that period for too long because uh the longer you are in the honeymoon, the more goes your anxiety that what you're

13:48 building is actually something that nobody wants. Yeah. Right. And so you have to launch eventually to to get that that feedback cycle or the flywheel uh how Amazon calls it. Um uh and so both of these things are true, but you're right. You know, if if if you are able to fund raise um based on reputation or based on on the idea you have and so the money question gets out of the picture, then you have a certain period where things are feel really good uh until you have to launch and and then you you're you're faced with the with the reality.

14:17 Yeah. with the reality of the app. So,

The software development lifecycle is undergoing foundational change

14:19 how uh I guess what can you say about your sort of uh new startup uh on it and maybe like again what made you sort of take the leap from from uh from where you kind of were uh in GitHub to this world.

14:31 Yeah, I mean you know what made me take the leap kind of explains where we're going. Um look, you know, I started um uh in be becoming a um company founder in the developer tools space uh back in 2009. Uh uh I left my my job at at Bosch uh uh in Germany, one of the big automotive suppliers. Um uh building if you have a car, you know, with ultrasonic sensors in the in the bumper, those are most likely from Bosch or one two competitors.

14:59 And uh I left because I was so excited about the the iPhone SDK and the app store and I wanted to become, you know, an app developer again after being in automotive for for six, seven years. And from being an app developer, I became, you know, the startup founder of a company called Hockey App that provided a developer platform uh uh for mobile app developers, you know, distributing a beta apps u collecting crash reports and feedback. And that company got sold uh or we sold to that company in in uh in late 2014 uh to to Microsoft. So it's been now 11 years uh since I was last startup founder. But ever since you know my hockey app days, I have been in developer tools. you know, hockey app was a developer tool. Then I was at Microsoft um together with NAD building app center which was kind of like bringing the pieces of Zamarind test cloud and Zamarind insights together with hockey app and we bought hockey uh then we bought GitHub uh you know I was uh VP of special projects at GitHub before I became the CEO and through that journey you know from mobile to the cloud um to co-pilot and AI I have seen

16:00 how how software developers have have changed the world and that the really exciting thing I think is about being a software developer is that as you're building new products, you're constantly also evolving how you do that. Um, and you know, if I remember my first coding uh uh uh exercises in the early 90s uh first on a East German robot uh uh computer and then on a Commodore 64 to what software development today is like it's it's fundamentally different. Like if I would give my kids a Commodore 64, they would probably, you know, figure out how to play Diana Sisters or or or something like that. but like typing basic and it's like what the hell is this user interface and where's my you know where's my my copilot or my stack overflow or or just Google to to look things things up

16:45 the internet exactly the internet you know the internet was open once you started doing it you couldn't go back to sort of developing without it

16:52 correct right and so if we we have always had these moments where the internet changed everything open source changed everything mobile changed everything the cloud changed everything AI is also changing everything I actually think you know we are still at the beginning of that revolution. Now, co-pilot started um more than four years ago. You know, chat GPT came more than three years ago. Last year was all about chat and and agent mode and and this year clearly it's all about full scale coding agent, code review agent, security agents. Uh but I still think you know we are at the beginning of a journey where the way the software development life cycle works is going to fundamentally change and that's you know what I'm excited about and that's what we're going to explore in my new company.

17:31 Yeah, cool. Yeah, there's so much I think the there's always like an underestimating kind of a kind of natural reaction, you know, when you think about uh just sort of the scale of the transformation, the

17:43 what's the sort of the uh the cliche is that, you know, people uh overestimate the short term and underestimate the long term, right? In terms of the potential of it. So, uh your survey,

17:53 it's really hard to predict, right? Like if if we go back a year um uh into December uh 2024, who would have predicted that at the end of 2025 developers are back in the terminal,

18:05 you know, with with Cloud Code and and Gemini CLI and Codex CLI and Coupile CLI and have like six or eight or whatever a grid of of terminals open and have all these agents running in parallel and trying to figure out, you know, uh how to make that work with Git work trees and how to merge all the results and how to not have the agent delete all the files on your computer or or the work of of you know the last uh 30 commits what have you and I think you know that's where if anyone who's predicting what that looks like in a year or two is going to be fundamentally wrong and I think that that's really what makes that so exciting we don't know where we're going other than that we do know that we're going to keep building software while maintaining you know billions if not trillions of lines of code that are already out there

18:47 yeah and that are growing and also like you know kind of a part of the challenge with the the tools so so there is like I'm still reading though like there's a fundamental conviction for like when you make this move that a lot of the ability to explore unencumbered is just sort of better to do in in the dev space and I guess from from a like uh it's interesting to sort of see GitHub like you know because as we as when I think about innovation you know from incumbents on it copilot is actually like GitHub copilot is a very kind of is a great example of innovation that came from sort of within the company uh and kind of like pioneered that sort of new space of of of coding assistant that feels like an AI tools now sort of you know like the the old pattern but you know for a while was there um and then like even with all of all of its reach and all of its sort of strength there was uh like the opportunity for cursor I guess kind of most notably to you know copilot continue to thrive so it wasn't really like at the expense of sort of copilot but but but from a like a market share perspective you know kind of grow even faster right and sort of take up

19:48 that sort of space on it and then indeed you have the disruption coming from anthropic which is kind hard to think about them. Are they a startup? Are they incumbents? Like th those are very weird beasts. The sort of the open AI and anthropic that are like especially anthropic like sort of a short-lived organization and a massively valued uh organization uh but still very much a startup in uh in many ways. Um and you know overhauls like the the fundamentals of it's going to be terminal versus ID.

20:14 Um, so I guess there's a there there's a sort of a a faith or still a perspective that when you're looking to disrupt like I I I believe that especially when you're disrupting the way things are done, not not just automating existing workflows but rather sort of changing the way that someone work those are precisely the places where where it's hard for an incumbent to kind of accept it because they have all the power in the existing way of working. uh and so they you know it's sort of it's hard to accept that that change will come right

20:46 I think that's the that's kind of the outside in perspective and that you look at you know companies that have been leading uh for a long time and then argue being disrupted or seeing more competition and you you know feel sad or worry about the people that are now kind like in that position that they are no longer the best in in the market. Um, I think the inside out perspective is often uh not as um dire as as you may think. You know, um I I've often thought when whenever for Mercedes uh that the best thing that can happen to us is that BMW announces a new model with a with a new thing in it uh that the Mercedes leadership didn't want to fund. And all of a sudden you have an argument to get to get the funding. Yeah. And and be able to say, "Okay, look, our competitors are doing this. we better wake up and um and and and also you know try to catch up again and you you saw in the news that you know Sam Alman declared a code red and Google had a code red and I'm sure you know SA had um or I not I'm sure I know SA had a bunch

21:46 of codes um over over his his tenure and um and so competition is good it actually makes uh business and it makes life more fun and we see that everywhere you know coming back to my Formula 1 example you know it was Red Bull that won for four years and it was Mercedes for like eight years and and that bullhead had another four years. You you kind of want these changes and it motivates those that have been winning for too long, you know, to come back and then see how do we now transform our organization to be a winner again. And if you you mentioned cursor and VS code, you know, actually the IDE space has been phenomenally stable over the last decade. Like I remember when I you know started on a Mac I think was 2008 or 2007. I started on Textmate and and Textmate was the revolution back then and and then as text made kind of like became stale and and text made 2 took way too long switch to sublime edit h and then VS code came in VS Code for the last decade was the default uh editor IDE for almost everybody. Um there's obviously you know some EMAC users and some Vim users and and what have you and

22:47 now we have you know real competition again and it's not only cursor and VS code with with copilot uh vinsurf is obviously in the play now recently uh Google anti-gravity came out uh there's a kilo code and ko and zed and and so there's a hu competition and that means somebody sees an opportunity there to to disrupt and to innovate and and ultimately build a a sustainable business and I think that really is exciting and not something to to worry about both as the incumbent and as a disruptor.

23:17 Yeah. Yeah. No, I I I agree and I'm not sort of worried about it. In fact, I I like the changes. I think like a uh it all of these require a moment of disruption that sort of you know makes sense that is that is hard for sort of the incumbent to follow. And you're right that it's always like competition is the the greatest uh motivator uh for it and the opportunity to drive. And at GitHub, you know, it pushed us last year in 2024 um uh uh to go from just provide having one default model in copilot to offering multimodel choice um uh where you know at at GitHub universe in in November October 2024 I announced that we now back then you know support uh claude oh what was even the version number cloud3 I think um and Gemini 1.5 Pro and and in the meantime if if you open cop um it has like a whole list of uh uh u major models and then you can bring your own model and connect open router and all these kind of things right so you can kind of see the direct impact of competition in the market leading leading to a better product uh for the incumbent as well

24:18 yeah yeah that's actually really interesting because I think one of the big questions that is on GitHub we don't have to sort of over rotate here on GitHub but uh is is that you know GitHub has kind of managed to be Switzerland in many ways you know for a long time and it's like hey we're going to work with all these IDEs when all you know there was VS Code but it's not a money maker it's uh it's a sort of like a free service to uh to the ecosystem. uh and now with copi it's like a little bit different like Microsoft you know as a whole makes a lot of money from actually being the sort of the provider of these different models right and sort of and and not the builder necessarily but the sort of the uh the executor a lot of the inference you know running on Microsoft is a is a substantial uh revenue source uh and uh and is invested and so that sort of question on sort of uh neutrality I guess or like you know do you want to sort of play nice with everyone as long as you use my agent uh is a is an interesting it's not entirely new like it existed in sort of GitHub in other ways and like in the apps in the sneak you know for instance like in the app security as an example

25:18 um but it feels stronger now in terms of that internal conflict uh to uh

25:24 well you know Switzerland Switzerland uh also still wants to make money and all the companies there and the countries none of this is like critical it's just a change right it's just a change

25:34 if you look at Microsoft um it is um probably the uh uh biggest platform company that is out there like it's not only copilot and and GitHub that have some form of you know neutrality offering offering choice it's it's also if you look for example on into Azure AI foundry uh which has like a huge range of models available so it's not just the open AAI and the the Microsoft develop models it's you know mistrial models and black forest labs models and lots of open source alternatives and you can deploy your own model right so in that sense um maybe the Switzerland analy isn't as good as just saying uh it's a platform company and platform companies by design want to have as many players on top of that platform you know which is GPU inference as an example or or you know CPU comput storage or what have you in Azure um because that helps both you know the platform provider as in Microsoft and you know the partners all to thrive

26:28 right I think that's really the the leverage um that Microsoft has as a business is the scale of of that platform and that's the opportunity for every startup to become I'm part of you know Microsoft startup program and I should ask you have you actually joined the Microsoft startup program and

26:43 we're in the program but we're we're sort of like in all the programs right you know where uh because they all commit until until you get

26:50 but you know at some point every startup gets kind of especially every B2B startup maybe less in B2C but B2B every B2B startups comes to the point where they either want to leverage Microsoft's distribution channel and an amazing enterprise sales team or or Amazon's or Google uh because the scale of these sales teams around the world and their skill set is is so far ahead of what a small startup can ever deliver without going through the pain of you know building out subsidiaries and building out local offices and local meetups and and and you know there's a lot of human investment that has to happen that you can't replace with AI because enterprise sales fundamentally is still a you know relationship uh relationship management exercise

27:32 super valuable and the marketplaces and sort of all of that and sort of access to it I I think uh let me kind of veer us off. You know there's like a world that you and I can kind of geek out on the business side. A little sort of make us a little bit more technical here uh on it. Uh and you know we were just sort of established that it's hard to predict but still let's sort of make some predictions here. One of the things that I love your opinion on is is kind of like the scope of products in in the era of sort of AI dev tools uh of it. Uh and maybe uh the the the first one is is just like the the delineation between sort of reviews and kind of you know uh collaborative interfaces maybe like the GitHub interfaces. I guess we're sort of still there but uh uh versus kind of earlier on premise right we see uh uh you can use agents locally you can use agents that are sort of background agents sort of the devins of the world that will you know now there's a whole variety of others uh that will that you'll always kind of meet the code at the PR uh to to begin with um and then

28:33 you see all the review agents you know I had Merryill from Graphite kind of on on the show here and you know we talked about how, you know, they they're adding, and that's true for everybody, an ability to now interact, you know, with the uh with the um uh the whatever, like not just find a problem, but also kind of fix it right there on the spot. And there's the question, of course, well, if you could find a problem at that point, couldn't you have found it uh beforehand on it? Yeah.

28:59 Um and a lot of the lines blur, you know, a lot of the you know, where does uh because automation is is much more increased. And I'm curious like what's your what's your lens when you think about the the um the division of products even between local devs and sort of reviews and collaborative.

Agent-to-agent workflows reduce human review bottlenecks

29:18 How do you how do you picture these things uh evolving?

29:23 We always picture those things as somehow collapsing into just one you know unified stack or suite of products or what have you. And I think the reality then is that never actually happens. And it's a bit of a pipe dream uh of every you know head of engineering or or engineering manager that they have just this beautiful stack that everybody in the company is using and and so you don't have to uh worry about you know outdated products or like conflicts between one product does this and one product does that you're buying multiple times. So I think let's be real like the you know there will always be uh you know overlaps in in what companies do and products do and there isn't like this oneall solution. Um sometimes it's you know purely based out of necessity of you know moving from one tool to another uh is is is is painful and you rather not invest that time and the literally the time and the money of your engineers doing that just because you want something completely standardized.

30:19 uh you're just accepting to some certain degree I think to c to large degree actually in engineering we will have to uh accept that entropy always exists in software and your software

30:30 and your stack always gets worse until you invest time to remove tech debt. Um, one of my jokes, you know, as as GitHub CEO has been always that uh uh debt is actually the wrong term because debt is something you're supposed to pay off over time. In in tech, we just refinance into into the next big thing and then there's always the next big thing and before you know it, you have new debt uh even though you just refactored from the old debt into the new debt, right? So, it's more of a refinancing.

30:55 In the business world, the word for debt is leverage, right? You know, it's just a different lens over there. It's it's you know it's it's it's it's part of you know it's like similar to uh your house in your kitchen that you never get you never get out of the mode of of putting the dishes into the dishwasher and out of the dishwasher unless you stop eating and then and and then then you die and that's actually true tool in software too unless you stop you know uh building new features into into software you're always going to be in a state where you have to you clean up um afterwards and and so if coming back to your agent questions I think you're going to see you know more of a you know agentto agent collaboration

2 Highlight · Moving toward agent-to-agent collaboration

31:30 Um so you mentioned graphite merl and and you know the code review agent what's really interesting about that right now is it gives me as a human feedback on code now but if that code was written by Devon like why do you give that feedback to me give it to Devon right and have Devon figure out like that that's how you would do that in a in a in a humanto human developer team right you would tell the two engineers to work together until they have reach stable state uh and then they can come to me as kind of like the manager of of of of these two humans or the two agents to say, "Hey, we have we're finished with our work." So, I think there's going to be uh um you know, you know, more we're going to see more of that uh uh like the agents working with each other uh to not make the human the bottleneck uh within the life cycle. Um, Simon Wilson on his on his blog recently had like a statement, I'm paraphrasing, um, but this was like, we have to get comfortable with not reviewing every single line of code before we deploy them into production because the sheer amount of code that a single agent can write 24/7. Now, you

32:32 paralyze that into running 10 agents in parallel all the time. you're just never going to be able to review all that code without the one human doing that becoming the ma major bottleneck and effectively erasing some of the productivity gains from these agents.

32:46 And so you have to think about how do you automate other processes down the line to to be able to generate all that code in the beginning.

32:54 I think one thing that's also like underappreciated is that a lot of these coding agents are actually uh really good at brainstorming. Um and so a lot of you know my developers are using cloud code to write code but before they do that they also use cloud code to brainstorm uh uh with the agent um you know to learn about a topic or about you know some special API or what have you or even implementation options right like even a different

33:20 what are the other options yeah uh to write you know an architecture document um uh while you know everybody else in in Australia is asleep and then and when they when they can up you already have something ready to share and so then they can pick it up and and and and keep working on this. I think there's and again this is very human centric right brainstorming you have to tell the agent what you want to work on and I don't see that going to be replaced anytime soon like the idea still has to come from the from the you know product manager or engineer or you know founder um and the how is is a iterative process with the agent but it's not going to be replaced by an agent making up the idea telling another agent to implement it and a third agent to review and deploy it

34:03 the agent to agent conversation is is uh is one I I relate to and I think it's still interesting when you think about the dev tooling ecosystem to think about who who does this right because you can run claude code as all of these agents right you can run you know any of these whatever codecs right like you can run that in the mode of you know like I wrote the code and run it with another persona in fact sub agents kind of do that right like you are creating you're you're delegating work or you're passing along to some agent that has the right context to represent expertise you could choose different models on it. Uh but uh but if they're if the agent is a is a human, you know, often times like from a from a company's perspective on it, it might be one company, you know, that is, you know, providing with the agent and you're connecting it in all these different contexts. Um so I guess one thing I still so so I I relate to and I'm curious about I don't know that we precisely know what the flow is. What are the interface points for us to uh to engage with larger bodies of volume of work? like review as a concept will

35:04 probably increase in importance because you do less of the work. you review the work of uh of AI but but a lot of the current like software development review activities maybe will go away and that's not it's not just for that right it's also like for testing it's also for you know like I'm I'm just picking on that as an example um and so I'm kind of curious and almost like a little bit concerned about a collapse of a lot of these different tools and and the reason I'm concerned about it is is your comment before about competition right is because you want you you don't want it to just be the competition between five giants, right? You want you want a composable ecosystem, right? And in which different uh pieces can kick in.

35:46 Um I don't know if it's like a comment or a question that I'm sort of making here, right? But I uh I guess you know a little bit of like what are the right delineations, you know, for for a company.

35:58 It's more if you think about it, it's more composable than ever though, right? um like nothing stopping you um to use cloud code today and Gemini CLI tomorrow and and codeex on Sunday.

36:08 Yeah, very um there's there's almost no switching cost. Same for the IDE uh same for the code review agent. Um uh and and again this goes back to um you know the platform concept like the platform isn't changing you're storing um your code on GitHub or or GitLab or or Bitbucket um or you know one of the you know few alternatives that out there um but on the agent side on the AI side everything is composable and the space is moving so fast that there's you know as we're recording this yesterday GPT 5.2 2 came out. Uh, and until that point in time, everybody, you know, was talking about Gemini 3 Pro being the best model. Now we're kind of like figuring out is 5.2 now the best model. And by the time maybe this part comes out, there's yet another model and everybody's like, what what the hell is Thomas and Guy talking about about what the best model is? And and then so as such, you know, there's like almost no switching cost. And that makes really this idea of I can compose my pipeline uh on the best tools I have available um so much more you know uh

37:11 realizable than than it was in the past when you did have uh vendor login right like if you built um your front end on react uh there's really high switching cost um to to go to a different uh front-end framework when you're building uh your CI/CD on on GitHub actions there's really high switching cost moving that away from to another CI/CD provider so in many other parts of the developer tool chain, we have high switching costs, but in AI agents is really really low and and that really creates this competitive envir environment where what was the best thing six months ago may not be the best thing now and whatever you know the best thing is in six months nobody can predict.

37:48 Yeah. I guess my my um I guess I would challenge a little bit like you had this sort of comment over there that like the GitHub and GitLabs and Bitbuckets like are still the place your store and like it's just the agent that you play around with. But I guess I would say that the platforms are optimized for a type of work. Uh and so you know the whatever sort of uh infrastructure estate management tools that existed pre-cloud were not the right tools to manage a cloud estate because in the cloud estate software or like hardware was elastic.

38:15 you know, you could kind of spin things out and and down and uh you know, identity therefore of these machines was different because you know, they were short-lived and you know, immutable infrastructure and all that. And you know, now we have uh a different breed of developers. The collaboration, you know, as you pointed out, right, happens between agents. And so what what type of uh um uh tracking you know and it it's hard to not think about git as that because git is uh is like we're so it's so ingrained you know and but I think we're both old enough to have kind of lived through subversion and sort of you know like you know a bunch of this sort of you know uh uh the places which sort of get felt like a heresy and um and but uh but you know fundamentally you're you are uh uh versioning and storing and collaborating ing uh on uh you know between different methodologies right on different competencies. Um, and I find it actually really curious to think about what is what is the right way to version in and and and should it be and

39:16 maybe this is a part of where my sort of pull request review I was like is pull request even like the thing to to represent anymore uh because you know is that really the point of collaboration uh between agents uh or or is there something quite different right? I think the question, you know, if we think about these collaboration platforms though, is do I still have to go there to figure out what's the work that's assigned to me, you know, what's my team working on or do I head into a world where I just ask my agent what's next, you know, on my backlog or what's what's guy working on? And so you're heading into this world where you know the

39:54 your your your agent um or your assistant if you will um becomes like you know Javas to iron man and effectively knows everything and and has sub agents you know to access certain skills. Yeah. And so you would basically have your personal you know agent that helps with travel booking and calendaring and and has you know access to your coding environment and all of a sudden you don't actually open the browser as much anymore and open all these tabs and and all these platforms. um where you know your your leadership team told you this is where your stack is um and that's why you have to look for work um

40:28 and so I think this is you know but at the same time there are still people running cobalt for mainframes

40:35 and you know most of these cobalt projects have no devops whatsoever they don't use git you know they they don't have CI/CD

40:42 they certainly have no sneak uh or or test cases and test suites and and what have you and and if you take that as the most extreme example you'll realize that as software developers we live on that spectrum where on there's the really old stuff that still powers you know huge part of the banking industry of you know the local and and and federal governments and so on uh and in between you have like you know people working on Windows APIs uh for displays on on on on train platforms uh uh Windows 95 APIs not not Windows 11 API

41:13 but those are and and so you have that spectrum and so yeah there is going to be the kind of like state-of-the-art developer that uses all these agents and doesn't have to go uh into the web anymore because all they can do is kind of like talk into the computer like Mr. Data and Star Trek and and gets the job done. But in between now all the other people that still need everything else that we built in the last uh uh in the last 30 years and and as such you know that that is is really like a a gradual change that's happening and it's not this oh AI is writing 90% of all code now and and we don't we don't need all this old stuff anymore. I think is now as as long as we are aware of when we talking about how exciting these agents are that in the middle there's like a huge spectrum um of of other work that still exists that engineers have to work on. Uh I think you know it's both exciting and also grounding uh uh how how much work there is going to be for for the next uh you know century to come.

42:05 Yeah. Yeah. For sure. And and that sounds very aligned with the sort of it's more of a decade of the agent versus the year of the agent, you know, from Karpath's kind of quote on it. And and it's true and you know, I kind of sometimes say how like application security companies are like cockroaches, you know, they never die because because technology never dies and you kind of need to sort of secure all the way these, you know, back to the middle ages, right? You know, sort of uh uh you know, you need to uh you need to provide support for that and clearly none of the new companies wants to sort of go there, right? and the inevitable kind of growth of any enterprise software company but specifically you know for that make

42:37 you have to go historically right

42:40 I think you unintentionally made a bit of a cynical statement about your own company so you could leave because I uh you know first of all like I'm very happy for sneak to be sort of uh you know immortal and you know I'm uh happy with that but it's no it's also like you just sort of see I've been in I've been in ABSAC actually for over a decade before I started sneak and and sort of have seen all these companies that you know just they they came and they toned down but they never went away uh because uh because of those so so it's interesting and I agree with you and I think technology is very different I guess kind of one topic to to uh to that that I feel like this leads into is like as individual as sort of the humans in this whole kind of movie uh you know when the cloud came along uh uh you you know I guess you could kind of appreciate that yes there's historical ones and Maybe you made kind of a version of a choice of like I'm going to really specialize in this sort of mainframe stuff uh because it's going to be a big money maker and I'm going to suffer but I'm going to make a lot of

43:41 money and uh and I'll sort of specialize over there. I'm slightly cynical here or maybe a lot cynical uh or or I can I can go and immerse in the sort of new world of cloud and and and and there was like maybe a little bit more clarity at least at some point around the skill set that you need to therefore build uh to be a cloudnative developer, right? to sort of be you know a capable uh professional in this new era and and AI is a bit more murky than that. So like a bunch of these predictions are you know beyond their intellectual curiosity. I think of course for you and I their bets around as we build dev tooling companies you know you have to you have to have a theory of the future and then be very adaptable that but you have to have a theory of the future but then you know for for so many more people you know there's a question of like I'm a developer what would my job look like?

44:30 What would be my competency uh in the future? So I guess what what do you what what's your view on that right? If someone comes along and says so what should I what should I get great at you know what should I uh um what how do you

Continuous learning and problem decomposition define future developer skills

44:46 envision the job of a future AI native developer?

44:50 Ever since I went to university I felt like the you're actually never done uh with with learning as a software developer. um there there is no stable state and those that feel like they have reached stable state are ultimately are planning for you know their retirement either voluntary or involuntary um and sure you know there is stuff that hasn't changed much um uh like IBM mainframes and cobalt code but I think you know if you if you look into your company and if you look into the world out there you realize that that's not a growth opportunity and the team is getting smaller and smaller and and you start working on more boring tasks and so yeah you have stable state, you can be like a specialist and maybe you can even you know uh ask your boss for for higher salary uh because uh the the boss will have a hard find will have a hard time finding uh college grads or university hires that actually want to work on cobalt. Um and um but that that stable state ultimately is is is over the long run not stable. It's just a gradual decline until this this reaches zero or

45:50 like a you know marginal value and and as such the biggest skill um as a software developer as an engineer is that you know how to learn uh uh new stuff and and how to put this new uh information in relation to what you're currently working on because not everything is actually as exciting as it looks and and um there's also the danger you're jumping you know on on the new train um uh and and then you realize Oh my god, you know this is actually throwing us back two years. Like when Apple released, you know, the Swift language for the very first time, a lot of developers were excited. And then they realized moving their project from Objective C to Swift, it wasn't actually ready and it was slower and and the the downloads all of a sudden was blown up by like 100 megabytes or or whatnot. And and and then you know there's many such examples where just jumping on the new technology isn't the right decision either. And so having that skill, you know, to learn and then to to judge, you know, to apply your craft, um your

46:51 craftsmanship that you have learned over the last decade or or so and ultimately, you know, um taste to to say, okay, we're going this way and and and not that way. Um and then and I think this is crucial even more so now with agents is really what you could call engineering skills like taking a huge problem and of you know the huge problem isn't as huge as building Facebook but it's huge enough that you know uh you

3 Highlight · The enduring importance of engineering fundamentals

47:15 don't know how long it will take you to actually implement this and breaking it down into into smaller into smaller pieces

47:21 right all the way down to okay so now I know what method and and unit tests I have to have to write to actually implement the the next the next abstraction layer and I think that engineering skill of you know decomposing a big problem into into a small problem then deciding you know how to to split the work uh uh in your own day and then with your other team members that is even more crucial with AI because only if you know what you can offload to an agent and the agent can actually implement efficiently uh and what you have to do yourself then then you gain uh productivity from AI right like if you just try and error everything here agent try this okay bad try this again bad at some point, you know, you could just go into the CSS file and change the background color yourself, right? And there's like some oneline changes where it's like idiotic to use AI if you know exactly where to do that. And you might use AI to just see how the agent fails with that and you have a good you have a good post on X or um uh but the reality is the engineering skill set will be to know when I can offload my work uh to an

48:22 agent or or you know work with a co-orker and then have to do it myself. But how so so I mean I I absolutely

48:29 agree with that today. But then what I would challenge is, you know, if you're, you know, a kid coming out of sort of university right now, right? And and you don't know that is the right sort of domain to sort of really double down on investing in getting good at that, you know, and and sort of knowing how to get down and know which lines in the CSS to edit versus versus sort of spending that the opportunity cost, right, of that time learning how to maybe like uh convey and sharpen, you know, like what it is that you want to build, right? or uh maybe even building like product and sort of a uh and kind of a user sensibility appreciation. So, you know, kind of leveling up either towards the kind of a a faster uh sort of progression, I guess, for most people towards the architect skills or into the uh uh the more sort of a product definitions or maybe some sort of new nent uh uh profession. But I don't know like feel feels hard for me to sort of justify the time invested even even if today you do you do um uh you do benefit

49:32 from having that type of depth of skills. Or is it that the kids um whether they go to high school or to college actually learn how to use AI much faster uh than you know the senior or or staff developer that has already been doing that for for a decade or two. um you know if I look um into my kids in school they have to you know submit essays and uh funny enough you can actually read the essays of all the other kids as well and you clearly can see who used AI to write it which I actually think is is cool like kudos you know to those kids that are like maybe it was short before the deadline

50:05 and you know and then the teacher probably can also see that this was written with AI you know the the long M dash and and and whatnot and

50:13 but they don't have the engineering skills like they will embrace the AI faster they have the engineering So like should they lean into the engineering or

50:20 they're going next time they're going to review the output and and prompt the AI to build a version where the teacher cannot detect that it came from AI, right? And so as you as you I think in many ways we should actually start in very early like first grade or whatever to teach our kids how to use AI in to not cheat in a test but to actually iterate, you know, and and and modify the results and verify that the you know when it talks about whatever the what it actually says is is the right answer.

50:46 And so I think teaching kids that skill of how to leverage AI uh to the best possible outcome that is what you know we we're going to see a new generation um you know of AI native humans if you will that in the same way that you know if you look at uh what is it gen Z or Gen Alpha grown up they have grown up with smartphones and so for them using you know uh WhatsApp and and signal and and and chatting with their friends is is totally natural and while you and I when we grew up we had to you know make a landline phone call. Um, and and then you, you know, had the parents on the phone and you kind of like had to pass that first barrier of explaining to them why it's now a good time for you to talk with your buddy or, you know, even worse with your with your future girlfriend or whatever, right? And so like you you see how the next generation is adopting these techn new technology and

51:35 I think they have ultimately a better life. There's there's challenges as well, but like it's it's it's liberating and democratizing and I think the same will happen with AI and and everybody will be able to learn coding. uh even if their parents have no technical background even a school is bad at teaching it because they can just they can just uh leverage you know the the co-pilot or uh the chatpt that's always there for them

51:56 and you know what and also doesn't judge them and doesn't say you're stupid or doesn't say you you should have learned that three years ago right that's the other thing that's really liberating is that you can ask uh your AI assistant any quote unquote dumb question and it will keep answering until until you run out of patience it will never run out of patience

52:14 yeah I would little I want to sort of uh stress a bit about the sort of the the engineering skills like in in all of that narrative which I fully buy into uh they didn't necessarily need to know the underlying you know when you get an image from AI you don't need to know how to draw to be able to sort of guide it correctly you need to know what you want uh uh better uh and and so with engineering I feel like we get drawn into this but you need to know how to engineer that today that's true because you're they're limited but in a decade is that still true Um, and I I guess I I get drawn a little bit more to like I guess my biases are more about like know what you want, which is actually pretty darn hard, right? Like it's not a it's not an easy thing to answer and then know how to convey it and know how to verify that it's been done correctly.

52:58 Uh, and to an extent let go of some of your sort of coding chops to sort of critique the AI's work or like do not invest in improving those. Yeah,

53:08 I think we're already in that state where um you mentioned the image example um where you can easily also detect what is AI slop um and at some point you're just not going to go to that you know substack anymore or you know um listen to a presentation of all the uh AI all the images as as poorly generated um stuff and so that basically forces the creator to get better at doing that and in a way that's the new craft right of learning how to use AI to get the better output in for images that you know 10 years ago than the craft was to use Photoshop or use use pencil and pen. But of course, you know, you could have made your same argument uh whatever 20 years ago when Photoshop first came out that it replaces the the you know artist that that draws with pen and paper is something that's very easier quote unquote. uh except you know now there's experts in Photoshop that are so much better than you and I in Photoshop and and so there's still there's still a craft that exists and and that and and we're going to see the same with AI.

54:08 There's those that are you know creating amazing images uh with the help of AI that that we cannot create without you know learning how to do that. And if you go if you take that back to software, look, you know, the reality is um uh commercial software projects uh and their businesses ultimately have to generate a profit in the long run. H and ideally you have, you know, healthy margins um or growing margins at least um to to sustain that business. Yeah.

54:35 That's not going to work if you don't understand what's happening behind the scenes. And that's not going to work if you don't have engineering and and business skills, you know, to work with these agents to create, you know, ultimately a product that returns, you know, enough money to pay the salaries and and and a profit for the company. And and so then the question really becomes is who are the companies that are fast enough to learn these these engineering and business skills and who are the ones who think they can get away with you know just wipe coding everything and and not paying attention to this because those companies ultimately will will run out of money lacking competitiveness uh that they they will not survive in the market.

55:12 Yeah. No, I love that you have to know the insides to sort of innovate in it.

55:16 Um yeah like we can go on this for an hour but I think we're sort of running out of time on it. super excited to sort of see what you build uh on it. Thanks for all the insights here. If someone wants to sort of follow your journey, you know, when you uh share more about uh your uh you know, your sort of uh next evolutions of your company uh how how can they follow you?

55:35 They can follow me on X um I'm Ash um there and I'm also Ashtom or just my full name on LinkedIn. uh in the in the last few weeks have been a bit quiet as we're working in STS on the on the new startup but uh certainly more to come uh throughout 2026.

55:51 Cool. Well, really excited to see what comes out. Thanks again for sharing this and thanks for everyone for tuning in. I hope you join us for the next one.

55:58 Thank you so much, guy.