#270 Thomas Dohmke: GitHub CEO Reveals How AI Will Change Coding Forever

0:00Fuse, coding agent, all you do is write the prompt. The prompt is a longer description, so arguably you're writing natural language to describe the problem we're trying to solve. and then agent mode or coding agent writes all the code and so for certain scenarios it's 100 code written by ai but you still have to write you know in in english so arguably human language is becoming the universal programming language in the future the developer will have an orchestra of agents and the role will be to be the manager or the conductor of that orchestra and i think ultimately the biggest skill for developers other than understanding technology and the craft of engineering will be to decide, am I doing it myself or am I using an agent?

0:41And at what point do I have diminishing returns to assign these three lines of code to an agent instead of just doing it myself? Building multi-agent software is hard. Agent-to-agent and agent-to-tool communication is still the Wild West. How do you achieve accuracy and consistency in non-deterministic agentic apps? That's where agency comes in. A-G-N-T-C-Y. The agency is an open source collective building the internet of agents. And what's the internet of agents? It's a collaboration layer where AI agents can communicate, discover each other, and work across frameworks. For developers, PersistMean standardized agent discovery tools, seamless protocols for interagent communication, and modular components to compose and scale multi-agent workflows.

1:46Build with other engineers who care about high-quality multi-agent software. Visit agency.org and add your support. that's a-g-n-t-c-y dot o-r-g hey i'm thomas and i'd like to say i'm not only the gilbert ceo i also am a developer and i've started coding more than 30 years ago in the early 1990s uh first on an east german uh uh clone of a z80 um and then on on a commodore 64 and later a 386 pc um i've a developer you know since since my mid teens i guess um and through different parts of my journey i ultimately landed first at microsoft who acquired my startup in late 2014 and then i was part of the team that acquired github in in 2018.

2:44actually now today as we're recording this is june 4 2025. So we're exactly seven years after the announcement that Microsoft will acquire GitHub. And through that announcement and the acquisition, I joined GitHub and GitHub's leadership team in October 2018, and then became the CEO three and a half years ago. And so you were living where in East Germany before the wall came down? In Berlin. So I grew up in Berlin, you know, one of the suburbs that has all the prefab buildings called Marzahn or Marzahn in German. So I grew up there, you know, I had a, you know, was 11 years old in 1989 when the wall, you know, technically fell, quote unquote.

3:32And then, you know, as Germany got unified in 1990, I finished high school in 97 and started a technical university in Berlin in 98 and then moved down to Stuttgart to work for Mercedes. And so ever since I left Berlin in 2002 and haven't been back other than for, you know, a few business trips and to see the family. And you said you began coding on the East German clone. Was that before the wall came down? Yeah, I think my history is a little bit muddy on this. To my memory, we discovered in the geography lab, these Robotron computers, which were East German-made computers, but very similar to Commodore 64.

4:15You would start with a basic prompt and then load, I think it was cassette tape in school. You would load software with a basic command. And so we started coding on that. and then the community center in the suburb had these same computers and we started going to computer club. I don't know when, I think that was after the wall had fallen. Obviously, the suburbs and Berlin itself didn't change as rapidly as it now feels in hindsight. And so a lot of things were still following the East German system. I think the education system only switched in September 1991, given that the unification day was October, which was after the school, 1990, which was after the school had started.

5:02So there was another year from 1990 to 1991 where we were still in the East German education system, even though technically the country had already been unified. And then only in 1991 that we switched to the West German system. So GitHub, I mean, I think of GitHub still, and I'm not a coder, as a code repository. but it's become much more than that. Can you give us just a thumbnail history of GitHub's development? And then we'll talk about some of the new things that you guys are doing. I would argue GitHub was always more than just the repository. The repository is the foundation, the primitive, as we would say, that has enabled what you might now call modern software developer collaboration.

5:49but it essentially invented how developers work with each other and in October 2007 I think is when the founders of github Chris PJ and Tom started working on it a few years ago we celebrated that by creating a poster that has the first 10 or so commit messages on it and as you may imagine a lot of those were kind of funny because obviously the founders at that time didn't realize what they will be building and how important that will become for the world it officially launched early 2008 um depending on whether you count you know the preview or the the actual release somewhere between february and april a lot of the early users of github were in the ruby on rails community given that the github founders were coming out of the community in San Francisco.

6:45And it started really as repositories. The pull request wasn't actually there in the beginning, but it came very, very quickly. But it already had like, you know, it's hard to think about this now as something revolutionary. But in 2008, being able to see a repository in a browser interface, clicking on it, it was actually fast. It had a nice modern user interface. That alone, I think, created a lot of attraction for folks to say, I host my repo on GitHub. It's easy to set it up, it's easy to push my code, and then I can just go there and click through the file tree and see the history of the files and the commit messages and so on.

7:31So repositories was that first building block collaboration with pull requests that GitHub ultimately invented was another one before the pull request open source maintainers would send divs, you know, files with changes around and then apply those changes locally to merge them back into the main repository. And later came issues and wikis and all these things that developers need to collaborate. And today, I think what defines with GitHub is that whether you're an open source developer or a commercial developer where they're working on a startup or in a hobby project, all that software development process is very similar to each other.

8:13While we think about these as different worlds, they're actually not that different. And oftentimes, you know, the four things I described are all the same person. And just in different times of the week, you know, they're doing open source, they're doing hobby projects, they're working maybe on a big idea, and then they also have a job where they're using GitHub. Yeah, it's always fun to look at people's profiles and all the different things they're working on. Just for the kids out there, where did people put their code before GitHub? There's many ways to answer that question. And the funniest, I think, is that you would just put it in a folder on your computer, and then you have a v1, v2, or date and time stamp.

9:02And then you would realize three days later that you didn't actually do that. And then you lost a lot of the changes or you can't hold back to what you wanted to do. And I actually think this is even today is still pretty prevalent that people just do that. If they're early in career, if they're working on just some of their first steps in software development, a lot of the code is just stored locally. the origins of version history is version control systems is probably like a history podcast in itself there was something called cvs which was the most used version control system and i started coding and during university and then subversion came out and that was very popular and sourceforge was was kind of like the place where you put your subversion repository to share of the world and then git was developed by the linux kernel team linux torvalds and others and mercurial was the competitor um at the time and and clearly git you know has has won that race but in 2008 when github uh launched um that was still a question is it is it git or is it material but a lot of the early github users came from sourceforge and from subversion and even today we're still migrating customers of their old subversion repositories into Git and GitHub.

10:28Yeah. Again, for listeners that aren't coders, describe what a Git is and what GitHub did for GITs. In its most simplest form, I think the way to look at Git is that it stores changes to the text files that you're writing. And developers, obviously, as they're writing programming language they're writing files you know plain text files with the code in it and git stores the changes that you make to that file at the time when you commit into the repository yeah and so it's version control it's the history of all the changes and nowadays you know invert and google docs and elsewhere you can also go through the version history but that wasn't always the case and the the significant change of git compared to other version control systems that came before it is that you have all the repository on your local machine and it's a decentralized technology.

11:28So you can have a copy of the whole repository with all the branches and all the changes. I can have my own local copy. And GitHub offers us a place where we can bring those together and where I can push my changes and you can pull my changes. And we can both collaborate on the same repository with each other. And so GitHub, as the name implies, offered the central place for all these decentralized repos so we can share them with each other because as much as we like decentralized technology and like to talk about the decentralized world, the reality is you still need a place, a home, if you will, to go to.

12:06And we're calling GitHub the home of all developers because we believe that that's our aspiration, if you will, to build a home where every developer goes and likes to collaborate with others and meet with them and exchange code. So let's talk about some of the new features that you've developed, particularly you've been very active this year. What is top of the list for you? AI clearly is top of the list. Otherwise I wouldn't be on this podcast, I think. And it's interesting because we talked about Git and GitHub and Git is a technology that wasn't invented by the GitHub founders, it was invented by the Linux kernel team.

12:54And then the founders took that a few years, I think two years before GitHub, and then the founders took that technology and built GitHub. And the same happened to us again with Copilot, where a technology was invented by Google, the transformer model, and then taken by OpenAI to build back in 2020, GP3, and a version of that model fine-tuned called Codex, fine-tuned on open source code. And we at GitHub at the time, you know, got access to this model through the partnership between Microsoft and OpenAI and started playing with it and asked questions, you know, how write a method that detects prime numbers or sort, you know, an array.

13:37Typical kind of like lower entry-level coding questions for for students uh uh in university and um it was able to write these these code uh examples uh in in proper programming language and in fact it could do that in different languages and would not mix things up uh between you know some languages like have parentheses other have curly braces some have commas other have you know colons and whatnot and that you know led us to the you know, idea of building a product around this, which then became co-pilot. And in its earliest form, it was really simple. It was just auto-completion in the editor.

14:19So where developers write their code and where they're using Git, you know, to create the version history of that code, they would now get a suggestion from this codex model or in general, the large language model. And the suggestions wasn't just like the next word or the next few words. it was multiple lines of code. You know, a lot of developers, a lot of what developers do is writing test cases and writing what we call boilerplate code, things that you repeat over and over again, you know, making around the corner so your user interface looks nice and so on. And so it could predict that code and as such it kept developers in the flow, right?

14:58So instead of switching between your editor and the browser or the editor and documentation or the editor and asking, you know, somebody in your office or in 2020 i guess you know on on slack or teams uh copa gave you enough of an idea of what you could do next even with hallucinations and all that that you could just keep staying in the flow um and often this flow state feels really magical for software developers it's kind of like i have this idea and now all i all i need to do quote unquote is just to you know take this idea and write down all the code to build what I have in my head. And so five years later, we're still working on that.

15:38And, you know, there's still a lot of innovation happening in this code completion space of predicting what's the next change after the current change, what's the next change in adjacent tabs or just making the model better. But ultimately where we're heading into is agents and building agents that developers can use to write code. We call that the coding agent, to review code. We call that the code review agent, or to fix their security vulnerabilities, and we call that the autofix agent. Are all of these under the GitHub coding agent umbrella, or is the GitHub coding agent referred to one of those?

16:17It's all under the GitHub Copile umbrella. So the way to think about Copile is it has all these capabilities, and we're adding more. So just like a human developer, you know, the co-pilot as a developer learns more skills and gets better in each of them. And so co-pilot has a coding agent, a code review agent, an autofix agent. It has auto completions. It still has to chat, which is often quite useful. You know, I have two sons. They're both coding in Python. And so they're using co-pilot, you know, to fix their own bugs instead of, you know, coming into my office and asking me. and often chat is the better way for them to learn than just having the agent do it because they understand then you know from the description and the code example of how to make the change instead of just magic happening to them yeah actually how old are your sons and and are did you teach them to code or have they learned on their own or at school uh they are 13 and 10 um And it's a mix of all these things that you said.

17:22So we encouraged them to learn, especially during the pandemic. They did, I think typing was the tool they used to learn 10 finger typing. I think they can do it better than I actually, because I never properly learned it. And so it's more like eight finger typing or seven finger typing. And they use Scratch from MIT and Lego Mindstorms, which has a very similar interface as Scratch, like, you know, logic blocks and whatnot. They learned Python in school, although the eldest learned Python in middle school. And then the rest, you know, they taught themselves very similar to how I taught it myself in the 1990s, except, you know, I only had books and magazines and they have the whole range of the internet.

18:10And that's incredibly powerful, you know, even without AI, just go on to, have a problem, you know, you're trying to figure out how to do, I don't know, animated sprites in Python and Pygame, and you can just find a YouTube video for that. In fact, you know, my kids also taught themselves, you know, to solve the Rubik's Cube by just watching YouTube videos, right? And of course, there's a lot of trial and error, but that is very similar to how many of us learn coding. And then Copilot gives you this extra step where it can look at your code, explain it to you, not only in English, but actually also in German and many other languages, which my kids speak both languages, given that we immigrated in the United States more than 10 years ago.

18:56But even that, I think, is incredibly democratizing that as a seven, eight-year-old, if you want to learn coding, you can just ask it in your mother tongue and and it will explain to you okay this you need to open a file and then this is where you put the python code and this is how you run it and now you have a snake game and yeah i think kids are in many ways actually better at this than adults because they are they're it's so natural to them to just keep trying right yeah they just keep trying they just keep asking as a parent you quickly run out of patience or daytime. But kids will just keep asking Copilot or ChatGPT or any other AI tool, and they will not sit there and say, this is bad, or I could do that way better without the model.

19:48No, no, they will take the feedback and work with whatever Copilot gave them to make their thing work. Building multi-agent software is hard. Agent-to-agent and agent-to-tool communication is still the Wild West. How do you achieve accuracy and consistency in non-deterministic agentic apps? That's where agency comes in. A-G-N-T-C-Y. The agency is an open source collective building the Internet of Agents. And what's the Internet of Agents? It's a collaboration layer where AI agents can communicate, discover each other, and work across frameworks. For developers, this means standardized agent discovery tools, seamless protocols for interagent communication, and modular components to compose and scale multi-agent workflows.

20:52build with other engineers who care about high quality multi-agent software visit agency.org and add your support that's a-g-n-t-c-y dot o-r-g yeah a couple of questions about the co-pilot coding agent.

21:19I mean, it's a tool built for coders, but for non-coders coming into this world, is it reliable enough that a non-coder could open up the co-pilot coding agent, ask it to begin coding something with chat, And because, as I said, I'm not a coder myself. When I, when, what was it called back then? You know, ancient history is like three years ago. The first of these that came out, I can't even remember what they were called. But, you know, you ask it, ask to code something. you then copy it into Visual Studio or whatever you run it, you get an error you cop the error back and you go back and you just end up going down a wormhole and never get it resolved so where are we now with these agents?

22:33Is it to the point that a non-coder could start using them? Yeah, I mean, there wasn't even Visual Studio when I started. You had basic or on the PC, it was Turbo Pascal. And it would run Turbo Pascal that actually ran in DOS, not in Windows. And then to run the application, you would move Turbo Pascal on Unix back then it was VI. You would move that into the background, type the comment line command to run the thing or compile the thing. and you run it and you get an error message and you have to figure out, okay, so how do I bring back the editor? And you wouldn't even see the error message and the editor side by side.

23:14Those things were on separate planes, if you will. Although there wasn't windows, right? It was all today. I think the state of the art is and the answer to your question is both. So as a non, somebody who doesn't understand programming language, you can use tools like Copilot. And, you know, maybe I should explain what the coding agent is, but you can use a Copilot. And there's a bunch of others out there, like Russell has V0 and Lovable Bold. Some of those, you know, do similar things, which is they can write, and actually OpenAI has the codecs now. You can write a prompt and then it, you know, creates all the code for you.

24:05And especially in web development, like create a quick web page or a small web application, even a little game, as long as it runs in the browser, you can easily create with these things. And you can keep going with more prompts in the same way that you can keep going, you know, when you want to render an image in chat GPT and you don't like the first version of that image, You keep going, writing more prompts, and eventually you either get there or you're like, okay, so I got so far off what I had in my head that it's easier to start from scratch. And I think that's between an image model or an image generator and a code generator.

24:42It's very similar. You can get a long way, but if you don't have knowledge of the programming image and how this all works behind the scenes, you're going to get into a place where at some point a prompt doesn't get you where you need to go. Right. And certainly when we think about scaling this then to millions or hundreds of millions of users, at some point, you certainly have to go and become a professional developer or hire those. Yeah. And the other piece is that, you know, and often this is called vibe coding, right? You're vibing by creating this, but you're always, you know, I think going to reach the point where now you have to learn what actually was generated by, by, by, by co-pilot.

25:30and to understand whether that code is actually, you know, doing what you described, you wanted to do. That's even more true in professional environments, right? Like when you join a company that builds software in 99 % of the cases, you're taking over somebody else's code and you're working or you're working in a project with, you know, that had dozens, if not hundreds of developers already on it. And so working on an existing code base and that's where the coding agent comes in. The idea of the coding agent is that you describe the changes that you want to make to the existing code base, and you assign Copilot to it, and then Copilot in the background, aka in the cloud, spins up a virtual machine, it checks out the code, it installs all the dependencies and all the tools it needs, and then it tries to figure out with the help of the model how it implements the issue or the task that you described.

26:26and as it does this it submits code into a pull request just like a human developer would and it describes you know how it made these changes and how it you know went through its plan but then you need to come in and say okay this is actually what i wanted and this fits you know into the coding practices of my team and it uses the right database and the right cloud provider and it doesn't introduce any new security vulnerabilities and that's where a professional developer software developer will always have to understand the code that the agent generates within the you know within the project they're working in yeah so you're sitting on the spectrum of if you want to do something without coding just to spin up you know your your wedding web page or like a quick e-commerce shop connected to shopify you know to sell something on on uh black friday that that definitely will work but there is a on that spectrum there is a point where you have to to then also understand the concepts of computer science, computer engineering, to ultimately be a professional software developer.

27:32And that profession isn't going anywhere. It will continue to exist and it's going up. It's not going away is, I guess, what I'm trying to say. Yeah, yeah. And what you describe, a software developer coming into work on an existing code base, does copilot coding agent how much of the code base does it review before it starts making suggestions or operating in a virtual machine in the cloud because Because, you know, there are so many, you change something, it could have implications that you don't see in the amount of code on your screen. Yeah. So does it review the entire code base or do you choose how much for it to see or how does that work?

28:33That's actually fascinating, you know, to observe for different types of issues they assign to the coding agent. And so it works in many ways, like a human developer would work in the code base. So it checks out the code base, it reads the description of what it needs to do. And then it feeds that into the model to generate the chain of thought. And as it does that, the model is allowed to call tools. And these tools can be things like using code search to find you know code within files to identify the files it needs to modify it can be tools that it runs on the command line uh you know to install something um or to update a package or maybe even add a package and what it does that comes out of the chain of thought that comes out of the model as it describes you know how it would solve that problem and but the key here is that it then looks at the output of these tool calls and, you know, trying to compile the thing or it can use the technology called MCP model context protocol that was invented by Anthropic that lets it effectively connect to every other system that already has an MCP server.

29:49There's one for GitHub and one for Figma, which is a design tool. And there's even one for, you know, can have it connect to your Apple notes or whatnot on your local machine. And so it effectively works like a human developer tries to figure out within the code base how to do the task. And then it iterates on that with the help of reasoning and the chain of thought. And until it ultimately gets to the point where what it has generated compiles and fulfills the test case, it created itself. and then it submits that for code review. But obviously, you know, that's never going to be perfect, right?

30:29There's always going to be either it thinks it has succeeded and you look at it like, what is this? And then you can give feedback. So we believe there's always going to be a feedback loop between the human developer that assigned the task to the coding agent and the agent itself. And that feedback loop at GitHub lives in the public best, just like the human collaboration between two developers lives also in the pool address. So if you go back in 2008, GitHub was human-to-human collaboration in software projects. Now in 2025, GitHub is still human-to-human collaboration, but we are extending it to human-to-agent collaboration.

31:09Yeah. And these models are getting more and more powerful. uh the the the underlying a model for github code pilot coding agent is uh are the open ai models is that right for the coding agent today um it is the uh claude sonnet 4 model um that just came out for copilot itself um the copilot that you use uh you know in your ide it is a range of models um code completions um or autocomplete is on an open ai model um that's based on the 4-0 mini um uh we also have you know for chat we have 4-1 and uh 4-5 and claude and google models and then for agent mode uh it's it's those models that uh support you know have reasoning capabilities and can call tools um you know at the end of the day whether it's the open source projects you're picking on GitHub, or whether that's the models, we believe in developer choice.

32:17A developer tool is not going to be successful if it doesn't offer developers the choice between the different options that are available in the market and let them pick which is the best model for their scenario. And what best means, and I'm making accurate for those that are listening audio, what that means is highly dependent on the developer themselves. They might have previous experience. They might have a team that says, we're only using this model or this open source library. They might have certain beliefs and all these things play a role. And we are not the ones judging that. We are the ones offering that choice.

32:56And that goes back to what I said earlier. We want to be the home of all developers around the world. And as such, we believe in that choice. So the developers can switch between models. presumably you know you can try it with you know doing the writing a piece of code with one model and then try it with another model and see which one you like better or but is there an orchestration layer where the co-pilot coding agent makes that decision for you so you know it decides on its own which is the best model to use? Today, the coding agent is only on Sonnet 4, but that's certainly something that we have in our minds or on the backlog, how we say in the software development, to have more orchestration on these things.