Transcript for "Agents in the Wild: What's New + What's Next": Hey there, everyone. Good morning, good afternoon, good evening, wherever you are in the world. Really excited to have you here today for this keeping up with copilot session. We'll be getting started in the next few minutes, so just hold on while we give people some time to join in. While you wait to get things started, it's a great time to get familiar with the chat and q and a options in the tool where you're watching this. So feel free to jump in there and say hello, and we're gonna ask you a a fun question, and we wanna hear all your opinions. So what's your it's it's been the Winter Olympics for the past couple of weeks. What's your favorite Winter Olympic sport? Or maybe just Olympic sport if you don't wanna choose an Olympic one. So give your answer in the chat, and we'll be getting started very soon. While I'm speaking during this during this webinar today, you're also gonna be able to use the q and a to ask questions. So jump into that q and a section to to ask questions, and we'll have some of the GitHub team answering those while the presentation while the webinar happens. And we'll also have a q and a section at the end where we'll do some questions live as well. Thank you so much for coming. Hey there again, everyone. So great to have you here today. Really excited to be having you joining us for our keeping up with copilot series. Today, I'm gonna be talking about agents in the wild, what's new and what's next. We're gonna get started in a couple of minutes, just leaving time for people to join the call. While you wait, I would love you to jump into the chat and tell us what's your favorite Winter Olympic sport or just Olympic sport if you prefer. Just jump into the chat and share what you're thinking. And also, while I'm speaking, you'll be able to use the q and a section as well to ask questions, and we'll have members of the GitHub team there ready to answer stuff and jump in and help out, as well as some live question time as well. We're gonna be starting in about a minute and a half, one minute. So hold on, we'll be ready to get going really, really soon. It's good to see there's a lot of variety out there. We don't have everyone just saying the same sport. We've got everything coming through. We've got hockey, snowboarding, figure skating, curling, all the different options. So clearly, you've been having a lot of fun over the past couple of weeks with the Winter Olympics. Yeah. What's my favorite Winter Olympic sport? That's a great question. I think the snowboarding is pretty cool. I always seems really, really hard and exciting. I also love the skateboarding in the Summer Olympics as well. Always always a lot of fun. I'm always just astounded by what people can do. I think if I if I jumped on a skateboard, I would just fall off immediately. So it's amazing what people can manage to do. Alright. Well, now it's about four past, so I'm gonna get started. Don't wanna keep you all waiting. Really, really happy to have you here. My name is Tim, and I'm on the product team at GitHub. I'm a principal product manager working on GitHub Copilot. I'm usually based in London in The UK, but today, I'm in Seattle in Washington, meeting up with a bunch of other people working for GitHub, building Copilot this week, which is really, really fun. We're always getting together to figure out the next things to move AI forward. And I'm really, really excited to have all of you here with us today. This is a webinar in our keeping up with Copilot series. And today, I'm gonna be talking about agents in the wild, what's new and what's next. And just to kinda give you a sense of what's coming up in this call, it's gonna be lots and lots of live demos, just trying to keep things exciting and fresh. I don't wanna talk through a thousand slides, but I wanna show you the cool stuff that we've been releasing and that developers at GitHub and elsewhere in the world are using to do great things. So lots of live demos, although there will be some slides as well. I'm gonna talk a bunch about stuff that we've released recently that's that's helping developers to do more and have more fun. I'm gonna talk about some of the stuff that we've got coming up next as well and share some tips for getting the most out of agents. And at the end, we're gonna have some time for q and a as well. Although you can also use the q and a while I'm speaking to ask questions, and members of the GitHub team are gonna be there to answer. So let's dive in. The reason we're all here today ultimately is because we're all makers and doers and builders. We're here because we love we love we love building stuff. We love building software. We love creating new things. I've been working in tech for nearly fifteen years now and and building software before that, you know, in my bedroom as a teenager. And building is what I love to do. And by nature, as people that are building software, like creating cool stuff, we're innovators, we're we're comfortable with change. We don't resist it. We're the people who are embracing change, and we're trying to create change. But all of you out here on this on this webinar, I have a confession to make, and I'm I'm just gonna put it out there and say it. My head is is really spinning from the past couple of months and the past couple of years. And that's because the way that we do software development has just been changing really, really quickly. That's not exactly new. Like, software development has always been a fast moving thing where things are changing all the time, but it really feels like the pace has picked up a lot in the last couple of years. When I was kind of first building software in the twenty tens, all we had to worry about was, you know, new JavaScript frameworks and new package managers and, you know, the pace of change wasn't that fast. But thanks to AI, we're seeing now every single week, there are new tools, new technologies, new paradigms, just new things going on all the time. There's new buzzwords we have to learn as well, new trends, you know. Over the past couple of weeks, we've been seeing a lot of love for OpenClaw or ClaudeBot or whatever it's called now. People super, super excited about these kind of agents. Remember a few months ago when context engineering was the thing that everyone was obsessed with? We Then had a period where it was all about spec driven development. So so much changing. Everything's moving so so quickly. And as always with this, there is hype and there are people trying to sell you things and sell you dreams. But at the same time, something really real is happening here and the way that we do software development is changing super, super fast. So let's trace through that timeline and and see where we've got to today. It's hard to believe that it's less than five years ago that GitHub Copilot first launched. That's just the kind of original version of GitHub Copilot that really started the developer AI revolution. I think I would say the AI revolution, full stop. That was the first thing that came out using models from OpenAI, so that when you're typing code in your editor, your Copilot could complete your sentences for you. It could complete your line of code. It could write your function and help you to move faster and and get more done. And that was the first step that began to change everything. And that was about, yeah, just under five years ago now. Then about three years ago or two and a half years ago, we had Copilot chat, a new way of interacting with AI that instead of, you know, just completing what I was writing, I could have a back and forth conversation. So stack overflow on steroids, my own personalized stack overflow to ask questions and get help. Or maybe even a pair programmer that I can get help from, you know. Instead of having to go across the desk to a friend or send a message to a colleague on Slack or Discord or Teams, instead, I can just ask Copilot and Copilot can help me out. And that's about two and a half years ago now. Then we had Copilot edits about a year ago, which went from just chat to Copilot being able to write code. And then just under just over a year ago now, we had the first agents coming out from GitHub. So agent mode in the IDE, where not only can Copilot now complete code for you or answer questions, but having AI agents that can write code and test and validate and iterate and get a whole lot of work done. And then we're in the middle of 2025, we had Copilot code review. So going from just the writing code part to helping you to review it and improve it as well. And if you look across that timeline, what we've seen is like the gaps between each of those things are getting smaller and smaller. So the gaps are the gaps are reducing. And we can see here this, like, increasing pace. Just things are moving really, really quickly, and the way that we do software development is is changing super fast. And this goes beyond just, the world of software developers, and it's it's leaking out into the real world as well. If you go to, like, any TV news broadcast or a newspaper, everyone's talking about AI and agents. And the first place they're looking to understand the future is what we're doing as developers. We're the people that are kind of leading everything. And and why do I think agents matter so much? Why do I think that it's such a big improvement or a big step change in what we can do as developers? Ultimately, it's because now we have AI that can't just answer questions. We have AI that can really do stuff, that can make progress on its own, that can be relatively autonomous. Of course, with us in the loop as well, but the AI is able to get a lot more done and really change our productivity, help us to do more, and hopefully take on the work that we don't like doing. That's what really excites me as a developer. Not just that AI can do stuff, but that I can hand over the things that I don't want to do to AI, so I can focus on the things that I find really, really interesting. And what I'm gonna be talking through today is GitHub's friendly crew of agents that are part of Copilot that you can use today. You've got our local agents, so agent mode in Versus code and the Copilot CLI. You've got our cloud agent offering, which we call Copilot coding agent. And you've got Copilot code review, our code review agent that can help you to review the code that you've written, whether that's written by a human or written by AI. So we're gonna be looking today across local cloud and code review. I'm gonna be doing a lot of live demos. There's not gonna be tons of slides because I want this to be super fun and, you know, seeing the cool stuff that you can do with AI and and how I'm using and how we're using it in GitHub. And I'm also gonna talk, as I said earlier, a bit about some tips of how we're how I found that we can use AI more successfully and get more done and have more impact. And also share some of the stuff that's coming up next from GitHub that I'm I'm really, really excited about. But before I dive into these different agents, I wanna start with some of the foundations of Copilot, because I think that's that's really interesting. One big thing that's shared across GitHub Copilot is this focus on developer choice. And we wanna make sure that whatever tools you like to use, however you like to do things, that Copilot's there and ready to help you. We're not, you know, forcing you to use one IDE or one set of tools or one model. But we wanna make sure that wherever you're working, you can be using Copilot and have access to the latest technology. And that's why we have Copilot available across Visual Studio Code, Visual Studio, the JetBrains editors, Eclipse, NeoVim, all these different places, and more as well. So we have Copilot integrations into tools like Raycast and Linear. So really trying to make sure that whatever tools you like to use, Copilot is there. That's something we've been doing for the past five years, and it's something that we're gonna keep doing because we think that's so, so important to give you that choice. Secondly, another big part of that developer choice story is really around making sure that you have access to the latest models from any provider. And right now, we have models on Copilot from Anthropic, OpenAI, and Google with more to come in the future. And we think that's super, super important, and we work really, really hard to make sure that the day a model comes out, you have it as a Copilot user. Because we don't want you to be, you know, locked away from the latest intelligence and the greatest things. We wanna make sure that when a new model comes out and it's amazing, you can try it, and you can experiment, and you have access to everything. We've released a lot of models in the past even two or three weeks. Think if I look at this, all the models on this screen released in the next month in the last month from memory, and there's gonna be lots more continuously. And I'm just super excited for how that's gonna continue to evolve because we've seen so much improvement in what we can do, not just from the products that we've built at GitHub, also from the models as well. So that's a really, really big foundation for Copilot that's strong today and is gonna be continuing to grow. Developer choice. Choice of what tools you wanna use and choice of what models you wanna use. Another really important foundation, which is maybe less important if you're a developer and more important if you're here as a kind of manager or leader, is really around metrics and security. So if you're giving AI tools to your developers, you wanna understand how they're using those tools and the impact that it's having. And Copilot metrics is a thread that runs through everything that we do. So making sure that whatever Copilot tools your team are using, you can understand and measure how they're being used so that you can track that and understand the impact that you're having. And similarly, we also believe that the kind of security and governance story is really important. You need to have control if you're an admin or you're a manager over what AI tools that your team can use. You need to have control over what models they can use so that you can use AI safely and you can understand how it's being used. So these foundations are across everything that we're doing, all the different products that we have. We're supporting all these areas across Copilot and making sure that whatever tools you're using, whatever agents you're using, you can benefit from from these useful things to help you. So that's the kind of intro part. And now I want to dive into actually looking at some of those agents that we talked about and showing you how I'm using them, how we're using the GitHub and and sharing some of the stuff that we've released recently to make these experiences even better. So first, we're gonna look at our local agents, and that's agent mode in the IDE. I'm gonna use Visual Studio Code, but we have these agents in other IDEs as well. And also the Copilot CLI, which takes your coding agent out of your editor and brings it into your terminal for a different experience. And we're gonna we're gonna look at both of those things. So let's get started with Versus Code as the AI native editor. Versus Code has been moving incredibly, incredibly quickly over the past few months and indeed over the past few years. On screen, you can just see me kind of scrolling through the change log for February, and there's just so many things being released almost every day. That's because we're just working so, so hard to bring you the latest, greatest things and help you to do more with Copilot. And part of the reason we're able to do that and we're able to move so quickly is the fit is the fact that we're using we're using Copilot to build Copilot. We're using AI to build everything that we do so we can move faster, we can ship more things, and we can get value to you quicker. Because what we love to do is help developers all over the world to do more, to be more productive, and to have more fun. So let's jump into the s code now, and I'm gonna show you some of the cool stuff that's been that's been going on recently. Obviously, now we get into the live demos part of this call. You can expect that some things are no doubt gonna go wrong. Live demos were always hard, and the demo gods, you know, don't always smile on us. But it's even harder when you're trying to do live demos with AI. So please bear with me if any weird things happen and stuff doesn't work, and we can just laugh hopefully and and have some fun together as well. So for the for the demos that I'm gonna do today, I'm gonna use a particular project that I've been working on, and let me just show that to you now. So this project is a kind of little arcade game that I call, guess the airline. So I'm a bit of a aviation nerd. Maybe some of you out there are there out there are as well, but I'm really into airlines and flying and all that kind of stuff. And if you've ever, you know if you've been in an airport, you've seen all the codes for airports, all the codes for airlines. And this is just a a little game that I built with Copilot where you have to oh, I think the screen share has disappeared. There we go. It's come back. So a little game where we see the code of an airline, and you have to guess what airline it is. It has a kind of fun arcade style. So here we can see that it's giving me the airline code DG, and I have to guess which of these airlines that is. And I don't know this one, so I'm just gonna guess. Maybe it's Sebago. Yep. I got it right. Great. That was a good guess. Let's go on to the next one. What's R K? Well, it's got an r in it. I wonder if it might be Red Wings. I know it's not Ryanair UK because I know what one that is. Let me try Red Wings. There we go. Well, I was wrong. My experience has been proven to be bad, and I've embarrassed myself in front of all of you. So there we go. That's the kind of that's the project that we're gonna use for this demo today. It's obviously very small and very simple, but it's a fun thing to build with and to and to get started with. So now you've seen the kind of basics of the project. Let me switch over to the s code, and then we can see some fun building with that. So I'm just gonna share my Versus code. And let me just make everything a bit bigger, so we can see it as well. So here I am in Versus Code. I've got this project open. It's built with a framework called ASTRO, which is just like a a nice simple JavaScript framework. It's a very simple app, so I pretty much got everything here in one file, to be honest. You know? One big HTML file or one big ASTRO file that's got some JavaScript. It's got some HTML. It's got some CSS. And everything really for the game is just in this file. It's super, super simple. Not built in a fancy, you know, beautifully architectured way. Just a really simple app. And obviously, on my right hand side, I've got Copilot as always. I open this by clicking the clicking the agent sessions button, the chat button in the top right hand corner, and I can see Copilot chat. And I've got three different ways that I can use Copilot today. I've got local, so I've got the kind of normal Versus code agent mode. I've got background, which lets me run multiple Copilot sessions at the same time, and actually under the hood, that uses the Copilot CLI. And I've got cloud, which uses Copilot coding agent. I'm gonna focus on local for now, but we'll come back to some of the other things later. One other thing to note is we also have our Claude integration in the IDE as well, which is pretty nice. So let's go into local. Obviously, as well here, I can choose models. I'm gonna stick with OPUS 4.6. That's my favorite model at the moment. But, obviously, other models are out there and depending on, like, the kind of work you do and your style and how you like your AI to talk and do stuff, you might like a different model. But today today, for this specific part of the demo, I'm gonna stick with Claude OPUS 4.6. So I've actually got a because I'm, you know, taking things seriously and trying to move my project forward in a nice way, I actually have a I actually have a Jira board where I'm tracking my project here and I've got my tasks that I wanna do. Obviously, different teams, different people using different stuff. Some of you might be using GitHub issues. Some of you might be using Linear. Some of you might be using Jira. But for this example, for this demo, I'm gonna use Jira. And I've got a couple of tickets here that I wanna get done that we're gonna work on together today. So let's click on this one first, says add fifty fifty lifeline to the game. And it's got a description of what we wanna do. So we wanna add a button that says fifty fifty that removes two of the options from the quiz, kinda like who wants to be a millionaire. So by default, I have four options to pick from. And if I click fifty fifty, it's gonna cut that down to two options. But I can only do this once per game. And, actually, this ticket's pretty good because it's already got a plan of what I'm gonna do. So let's take this ticket and give it to Copilot. I'm gonna copy the URL using this copy button, and then I'm gonna switch back to Versus Code. So let me just go and do that. I'm finding my Versus Code. There we go. And, actually, in Copilot chat, I'm just gonna say implement oh, actually, I need to install the Atlassian MCP service. Let's do that. So one of the great things here is that we've got MCP integrated into Versus Code, I can connect it to all sorts of other tools. And one of those tools is Jira. So I'm gonna open the command palette, and I'm gonna type add server. I'm gonna go to browse MCP servers, and then I'm gonna search for Atlassian here, then I'm gonna click install. It's gonna ask me to authenticate, which I'm gonna do on my other screen. Maybe it's gonna do it for me. There we go. It looked like it already did it. Remembered my session from last time. So now I can say implement this and copy and paste the address of my Jira ticket. This is pretty fun. I don't have to, you know, don't have to copy and paste all the context. Because of the MCP integration, Copilot can just go and grab context from Jira and use that. So it's just gonna go ahead and do this for me now. It's just figuring out what to do. It's realizing it needs to use the Jira MCP server, and it's gone to grab that plan. And now it can go ahead and do the implementation. So it says it's going to understand the existing code before implementing the changes. And now it's gonna get working on this. So it's gonna make the changes to the code in line with what's described in the Jira issue. So let me just open my ASTRO page, and then we can see what it's doing while it's working. There we go. It's added a fifty fifty button. That that CSS looks like it makes a bunch of sense. And this is where the agent magic comes in. So we see that it says starting, build, and test. So the agent isn't just, you know, writing code and then hoping that it's gonna work. It's not just YOLO ing it and making it my job to figure out if it did the right thing. It's trying to validate its work as it does things. So it's running the build to make sure the code actually compiles. And then afterwards, it's gonna be able to run my tests as well. There we go. Now it says it's gonna run the end to end tests because I have some playwright tests in here. Go to e2e and game.spec.ts. We've got some playwright tests. I'm gonna say allow, so let Copilot run playwright test. And then those tests are gonna get executed. Running nine tests. Nine out of nine. There we go. Nine passed. So not only has Copilot implemented this fifty fifty option for me, but it's also gonna allow it's also kind of validated its work as well and run the tests and build to make sure that this is actually gonna be successful. So let's jump back across to my browser, and we can see how the game looks now. Let me just share this. There we go. Now we've got our fifty fifty button. So I don't know this one, so I'm gonna click fifty fifty. It didn't work. Bad luck. Let me try refreshing. No luck this time. I should have asked it to write end to end test for this. So we've seen that that doesn't actually work. So let's Copilot to get let's get Copilot to fix it for us. So I'm gonna switch back across to Versus Code. Let's just tell Copilot, the fifty fifty button is there, but it doesn't actually work. Can you use the Playwright MCP tools to figure out what's going on and then get it fixed for me. And that works because I have the Playwright MCP server installed in Versus Code, which allows Copilot to open a web browser and interact with that browser, which we can then use to try to figure out why this isn't working. So it wants to run the app in the background, so I'm gonna allow it to do that. I think that's gonna fail because I was already running the server in my terminal. Maybe not. I'm not sure what Copilot's doing now. Maybe you can give us some encouragement. Oh, there we go. So now it says navigate to a URL with URL with Playwright. So let's say allow in this session. And then actually Copilot's opening up a browser. Let me share my screen a bit better so we can see that. If we switch across, gonna go to screen. I'm gonna say entire screen, which is always a bit risky, but we'll let it do it. There we go. So now we've got Versus Code. We've got my Playwright window. And now we're gonna let it navigate and try to figure out what's going on. That's the wrong window, think, actually. There we go. That's the playwright window that it's using. So it's good. I can see the page. Let me click the fifty fifty button and see what happens. So I'm gonna say allow, and then nothing happens. There we go. It found the issue. Use fifty fifty is not defined. So now it's gonna try to fix that. And I love these kind of live live debugging flows. It's so cool that, you know, these agents are not just, you know, writing code, running the build, even just running tests or writing tests, but they can actually interact with the browser for real and try to get things working and validate it for me. And there we go. Now we have a working fifty fifty feature that Copilot's figured out for me. It was able to look at what was happening in the browser, fix the problem, and get everything going. Super, super nice. Okay. So we've implemented that feature now. Let me stop sharing this screen. We can go back and do the next thing that I'm gonna do. Or maybe I can, actually. Let me find my let me find my Jira screen again, and we're just gonna go back to that quickly. So next Jira. This is a bit too small. So we've done our first ticket. Let's move on to the next one now. So this one says, remember score and game state when refreshing a page. And this is my favorite kind of Jira ticket. The kind of Jira ticket where there's no context, there's no description of what we wanna do, there's no requirements. So let's get Copilot to help us with that. I'm gonna copy the URL again, and I'm gonna jump back across to Versus Code. So let me click stop sharing and go to Versus Code. Alright. Make this a little bit bigger. Now we're gonna start a new session here. I'm gonna say I'm gonna say type slash plan to go into Versus Code's plan mode, and I'm gonna give it the URL of that ticket. And this plan mode is really nice because it prompts Copilot instead of jumping in and implementing something immediately, it's gonna help me to plan first and figure out what I what I actually need to do. So it's gonna go through the code, figure out what it wants to do, and then work with me on the requirements so that when it starts writing code, it's doing what I actually want it to do. So me and Copilot collaborate to figure out the right scope, the right changes, and then I leave Copilot to go and do them. So what it's doing right now is it's just reading through the code base a bit to understand what code we have right now. Then once it does that exploring, then it's gonna go and grab the Jira issue. Still busy exploring. Reading lots of files, figuring out what's going on so it can actually do the work that it needs to do. There we go. It's finished exploring now, and now I'm just waiting for it to go to Jira. There we go. Great. Here we go. I found the Jira issue. Let me read the actual game state codes closely to understand what needs to be persisted. Here, I can see that at the moment, only the high score is being persisted in local storage. But in my Jira ticket, I said that I want to store the whole game state. So remember where I am in the game and what I've got done. So remember the score one game state when refreshing the page. Still analyzing. And in a second, it's gonna come back to me with the plan of what it's gonna do. And I really love this flow because it allows me to make sure that Copilot is doing the right kind of things and, like, I'm actually getting the right code done rather than Copilot heading in a totally random wrong direction, and then I have to correct it later. Instead, I wanna use planning to figure out upfront the right stuff. There we go. Now I have all the context needed. Let me create a plan. And here we go. Here's Copilot's plan of what it wants to do. And I can read through this. I can take a look at the plan in more detail, and I can also press that start implementation button to get Copilot to jump in. So that's what I'm gonna show you from Versus Code today. Just jumping into the kind of agent experience here, the fact that we can have Copilot write code, run the tests, run the build, even open a browser and interact with it itself and try to use it with MCP servers, connect to Jira, all these kind of things. So we have this, you know, incredibly intelligent thing that can do great work for me. So let me jump across the CLI now because that's the next thing that I wanna show you. Let me press stop sharing. Maybe I've already done that. Let me go to my CLI window. Okay. It's now in my CLI. I'm gonna clear this, and I'm gonna open the Copilot CLI by running Copilot. I'm gonna use the banner org, though, because it shows a super nice shiny animation that I always really like to see. There we go. So the Copilot CLI is kind of like the agent that we have in the S code. But instead of running, you know, in my IDE, it's detached from that. It just runs in my terminal. Ultimately, the agents are very, very similar. They do actually have slightly different prompting, and they work in a bit of a different way. But they're all powered by Copilot. You can use them with your Copilot subscription, and they have the same models. But I think the magic thing about working in a CLI is that it maybe allows me to stay a little bit more detached from the code, so I'm not having to look at everything while it's doing things. I can easily run lots of these in parallel at the same time. So let's ask Copilot to add a fun feature. I'm gonna say, make it rain confetti when the user gets the right answer. I'm using Claude Opus 4.6 here again, and I've set it to the high reasoning mode because I wanna have the smartest agent I can get here. There we go. But what we're seeing here, this behaves very much like what we already saw in Versus Code. So Copilot can read files. It makes changes. It runs the build. It runs tests, all those kind of things. But instead of being in my editor, I'm just, you know, in a terminal window, and I can have multiple tabs and all that kind of stuff. There you go. So I can run multiple Copilots at the same time. Nothing's stopping me doing that. And as I said earlier, we also have an integration of the CLI into Versus Code as well. So if you pick background from the little picker in the chat window, you can have multiple agents running at the same time that are powered by the CLI, which is pretty nice. There we go. It started working on this, and it says it's added the confetti function. So let's actually give that a go and see if it's working. So let me run. I'm in the wrong directory. Let's see if Copilot gets it right a lot this time. Last time, we got unlucky and the first implementation didn't work, but I have faith. Let's refresh this. Alright. What's s g what's n j gonna be? Let's try this one. Nope. Oh, I'm getting hurt here. I have I have to actually get an answer right. Oh, that wasn't what I wanted. I have to get an answer right before I can before I can get any confetti. Usually, I get all these right. Oh, that wasn't quite what I had in mind for the confetti, but we can ask Copilot to fix that for us. So let's jump back across. And I can just keep having that interaction with Copilot. I can ask it to fix things. I can ask it to write tests. I can ask it to run tests, all those kind of things. We have a lot of other great stuff in the CLI that you should definitely explore as well. We have our slash plan command in the CLI as well, which is which is really nice, you know, just like what we saw in Versus Code. We shipped the slash research command earlier this week, which I really love. It can do deep research into your code base and answer questions and help you to understand architecture, understand how things are working, which can be really, really great for trying to build trying to understand before you build. And lots of other stuff as well. MCP support, all that stuff is in there as well. And again, like Versus Code, this is moving super, super, super fast. We're building so many things here. We're releasing new versions of CLI almost every single day. So really fast pace, lots of new things to try all the time. So make sure you stay on top of it and and try this if you haven't tried it already. If you're if you're using Versus Code, there's nothing wrong with that, but I definitely recommend trying CLI as well because I've loved it, and it's changed the way that I do development. Let's see. Did it get the confetti working now? I'll give it one more go, and then we'll move on to the next thing. So let me press guest the airline. Still not quite right. We'll get there, but let's move on. So now we've seen we've seen the Versus Code Copilot. We've seen agent mode there, and we've also seen we've also seen the CLI. Now I wanna show you our Cloud Agent, which is something that I've been working personally in and something that I'm really excited about. But before we jump into that, let me just show show you some slides quickly. Okay. So if I press this here we go. So Copilot Coding Agent, the idea is pretty simple. The idea is that Copilot works in the background while you focus on the most complex and high value work. And rather than running on your machine with Copilot Coding Agent, Copilot can work in the cloud in its own development environment. So if I shut my laptop or disconnect my Internet or any of those kind of things, Copilot isn't dependent on my machine, and it can keep working. And you can use this kind of Cloud Agent for all sorts of things, the same kind of stuff that you can do in in Versus Code. And the workflow of using a Cloud Agent looks like this. So you give Copilot a task. Copilot works in the background, and it it creates its own development environment that is powered by GitHub actions. And then once it's finished working, it requests review from you. And this has lots of integrations into lots of great tools. So it's integrating to GitHub.com and GitHub issues and GitHub mobile. It's also available from the IDE, so we saw that cloud option in the picker in Versus Code. You can send tasks to the cloud to the cloud from there. We've got integrations into Teams and Slack, you can just at mention Copilot to ask it to make changes when you're working with your colleagues or your team. And we have integrations into project management tools as well. So being able to assign an issue to Copilot or assign a ticket in Azure boards or linear to Copilot and a Jira integration that's coming soon. So then you'll just be able to click one button in Jira to assign a task to Copilot. Copilot Coding Agent has been around for about eight or nine months now, and it's had a lot of use. Back in October, we announced that Copilot had had over 1,000,000 pull requests merged, including lots of big companies. There's a lot more PRs than that. Unfortunately, I can't share a number, but lots of PRs. Huge numbers of developers using this to get a lot of work done. And your computer isn't even involved. It's all running in the cloud, and you can let it run-in the background. You can do what you wanna do. You can be in a meeting. You can be coding. You can be grabbing a coffee, and Copilot's doing its thing in the cloud. So let me jump in and show you a bit more about how that works so we can see it for real. I'm just gonna change my screen share again. We're gonna jump across to my GitHub tab. So here you can see GitHub, and this is my one of my accounts, and this is our guest, the airline repo. And what I can do here is I can just ask Copilot to make changes, and there's bunch of different ways that I can do that. So one way is that we have the agents tab in the repo. I click into that, and I see a list of the different agents that I've run, and I can just ask Copilot to do something. So I can say, for example, we saw at the beginning that the code for this app doesn't use React. It's just kind of a big file with lots of vanilla JavaScript. So let's say, continue to use Astro, but use React for components. I'm gonna use GPT 5.3 CodexFlat because I've been having some great experiences there, and I'm gonna click start. Not sure why it's doing that, but fix now. Let's head over to the pull request. So from that request that I did, Copilot has opened the pull request, and then I can jump into the logs. And I'm gonna be able to see Copilot start working on that in a few seconds once it starts up. There's a bunch of other ways that I can ask Copilot to do things as well. I can also use this agents panel that is available on every page on GitHub, and I can just choose a model, put in a prompt, and ask it to do something. So I'm gonna say, set up ESLint and run it in GitHub actions on push. And for this one, I'm gonna use Summit 4.6. And I press start. And there we go. Copilot's gonna start that one, and I can have many of these running at the same time in parallel. And let me show you one more option. I'm gonna jump into an issue that I created. And here, I can just assign this one to Copilot. So I can click assign. I can put in some extra instructions here if I want to, and then I'm gonna click assign again. And then Copilot's gonna start working on that as well. So then Copilot starts working in the background and does its thing. So let's jump into an example that I did earlier. This is why I asked Copilot to migrate to React, and it's opened a pull request for me, and then it requested review once it was done. So let's jump in, and we can see the file changes here. So Copilot's written a bunch of code for me, made all the changes, and I can review this and approve it. I also, just like in Versus Code or CLI, have access to the lock, so I can click view session. Then I can see how Copilot did the work. So I can see what files did it look at, what changes did it make. And just like running on my machine, it has a full command line available to it, so I can install dependencies, it can run tests, all those kind of things. You can go through here. It ran the end to end test, and it checked that the test passed. One last thing I wanna show you in here, which I really love, is that Copilot actually does a bunch of extra steps to validate its work. So once it finishes, it asks Copilot code review to give feedback on its work, and that allows it to do a better job before it hands over to me. It runs a CodeQL security scan as well to look for to look for any security vulnerabilities in the code that it's written, so it doesn't request review from me on code that's vulnerable. And it also is able to check dependencies. So when it installs a new package, it's gonna double check whether that package has any has any vulnerabilities, and it can tell Copilot to change that before it moves forward. This is really, really nice. I've had Copilot do stuff, I've I've handed over to Copilot, and it opens a PR and then texts me for review when it's done. We've got these other ones that I got started on earlier. So this one's jumping in now. Hasn't quite got started yet. Usually, takes about thirty seconds to start. There we go. This one's running. The one that we started a minute or so ago, and Copilot's getting to work and building what I asked it to build. This is all running on the cloud. So I can shut my laptop now. I can disconnect, and it's it's still gonna keep going. Alright. So let me jump back to the slides because I wanna show you a couple more things here before we hand over for some tips, some stuff that's coming up next, and some questions as well. So I'm just gonna get my PowerPoint running again. There we go. So let me tell you some of the ways that we're using Copilot Coding Agent at GitHub to help us to get more done. Because this kind of new Cloud Agent can be you know, it's different. We're used to doing things on our local machine. We're used to, like, having an agent in our editor, but having an agent in the cloud is quite different. So how are we using this? One way that we're using it at GitHub is to help us to ship new features faster. So we found that for kind of small and medium sized tasks, we can very often hand them over to Copilot to do in the cloud and get a really, really great result. So last year, we made some improvements to our secret protection product, which basically detects when you commit or push secrets in your code like API keys. We added a bunch of new validators to that product, which basically means that when you push something that looks like a secret, GitHub can check if it's a real secret, if it's valid, and then it can warn you and even revoke that secret for you. And we built a load of new integrations there to be able to detect more secrets and automatically revoke them. And we were able to speed that up really, really quickly. Previously, we'd had a really big backlog of these validators to check secrets, and we're able to burst through loads of these really quickly with Copilot. And we actually delivered about 19 new validators in just, I think it was a couple of months in the end, when before it had been taking us weeks for each validator we were doing. So really big performance improvement. Another opportunity we've seen with these Cloud Agents is allowing us to pay down technical debt without slowing down our feature development. You know, as developers, we often have this trade off. Like, we're trying to build new things and keep moving forward, but we also wanna make our code better because that helps us to move faster in the future. What we found is that Copilot is Copilot can allow us to focus on the new features while it does some of that, you know, refactoring, test coverage improvements, and things like that in the background. And we had a really nice example from the billing team at GitHub who used Copilot coding agent to increase their test coverage, and then that allowed them to move faster on the projects they were doing. And the final example that I wanna give is that these Cloud Agents let you code where the inspiration strikes. So now I can code even when I'm not on my computer. So when I'm going to grab a coffee or when I'm at the gym, if I have an idea of a new feature that I wanna build, I can just ask Copilot to work on I don't have to wait until I, you know, go home and have my laptop. I can just ask Copilot to do it for me, and it's gonna jump in and work on that. And then when I get back to my desk, I can review it and try it out. Or maybe even before I get back to my desk. Right? Like, if Copilot gets done really quickly, I can quickly review the code on my mobile, take a look, make sure that it's good, and then move forward. So here we go. We can see here Copilot has got this PR, and it's gonna work on this in the background and then tag me for review when it's done. And a final nice example I want to talk about is the ability to hand over your PRs to Copilot as well, which I found really, really useful recently. So the way that that works is that when I'm on GitHub, I can actually just at mention copilot in a PR that I created and ask it to make changes. So let me jump back across to my repo. Here we go. I'm gonna go to pull requests. And here's a pull request that I created earlier. As you can see, it's got a red x because the tests are failing. So I'm just gonna say here, at Copilot, fix the failing tests. I'm gonna go for some of 4.6. Leave a comment. And then not only can I get Copilot to start working on new things to, like, build new features for me or do refactors, but I can actually ask Copilot to work on something work on my PR and make it better? Right now, this is gonna open a PR on top of my PR, which helps to keep the changes isolated so Copilot doesn't, like, potentially mess up my work. But we're working on a change that's gonna ship really soon that will allow Copilot to push directly to your pull request, which is gonna be really, really nice and make it super fun to work with Copilot on github.com. So I'm really excited about that. One last thing that I wanna show you that we haven't talked about is custom agents, and this is saying that works across all the agents that we're looking at today. And this allows you to kind of customize Copilot's personality. This is another one of my repos, and I'm gonna go to the .github directory. And then in here, I have a directory called agents, and I've defined an agent called performance optimizer. And what this does is it kind of gives Copilot a special personality. So I can say when I'm doing performance optimizations, I want Copilot to take a data driven approach. I don't want it to just YOLO a performance optimization. I want it to, you know, actually measure the current state, then make improvements, and then understand how much better it got. And I can pick these custom agents in Code. I can also pick them in Copilot coding agent, and I can also pick them in the CLI. In the Cloud Agent, I just click this Copilot button, then I pick performance optimizer, and then it's gonna use that personality as it works. So let's look at an example of that and how it changes things. So I sent this prompt earlier and asked Copilot to to improve the performance of the of the agent. And if we go to this view session, we're gonna see that Copilot took a very methodical approach here that I asked it to take. So rather than just hoping to make things better, it actually is able to run benchmarks and measure how it did. Where was it? Hopefully, did. Usually, it does that. Yeah. So I ran a bunch of measurements to actually understand I'm making a performance improvement. Is this actually working or is it not making any difference? And then when I go back across the pull request, it's got information, so it can say it's improved the performance significantly. It made the start up take 21% less time, which is pretty cool. And Copilot measured that and can tell me about my impact. So that's custom agents. You can use them in a lot of different tools. So really cool and something that I'd recommend playing with. Alright. So that's most of the agents that I wanna show you today. The last thing I wanna talk about briefly is Copilot code review, which has been around for about a year and a half now, but has been getting much, much better recently. And I think Copilot code review is especially important now because with AI, we're seeing that the pace of development that we can do is speeding up quite a lot. We're able to build more and build more quickly, and then review becomes even more important for us. We need AI to help us do that faster and to get through the extra PRs that we're creating. And Copilot code review has got much, much better recently, and that's because it's also become more agentic. The first version of Copilot code review was, I guess, it's fair to say, we just showed it the code changes and then asked it, what do you think? Whereas now we're giving Copilot the ability to read other files, to maybe even run commands and stuff like that, and that allows it to give much more detailed feedback. So now I'm seeing when I use Copilot CodeView that it's spotting a lot more bugs, a lot more mistakes, things where maybe the description, the issue doesn't match what I actually built. And this is helping me to build much, much better codes. That's something that I'm really, really excited about. So that's the end of the stuff that I wanted to demo to you today. So what we've seen is Copilot's agents across local, so in my IDE and in the CLI. The Cloud Agent, we've talked a bit about code review as well. And these this crew of agents, this family of agents comes together to help me as a developer to be more productive and have more fun. And that's the thing that I'm most excited about. It's not just productivity, but helping us as developers to spend more time doing the fun stuff and less time doing the boring stuff. So if I can ask Copilot to do the boring refactors, increase my test coverage, you know, build GitHub actions workflows, which isn't always the mundane most fun thing, or maybe fix my failing tests, I can spend more time doing the stuff that I love to do. But let me share a couple of tips on winning with agents before we jump into questions. My first big tip is when you're working with agents, it's a great idea to plan first and then ask the agent to write the code. That helps us to get a better result because we have clear requirements, and then the agent can do exactly what I want rather than maybe producing something a little bit broken, and then I have to come back and fix it later. So plan first with an agent and then write code. The second thing that I'd really recommend is that you invest in how you set up your project to work with AI. There's a bunch of different ways you can do that. You can create custom instructions like agents dot m d files. You can create skills, which is another way that you can teach agents how to do things. And you can even create custom agents that we just looked at a minute ago. And we've seen that using these tools allows us to make AI way better and make these agents more effective in our projects. So just like you teach a developer on your team to to work better by having documentation, these are kind of like documentation for our agents to help them to be smarter. A third tip is that agents aren't just for coding. You can use them for lots of things. I, for example, use agents a lot for data analysis. So when I'm trying to understand, you know, how much Copilot is growing and how many people are using it and these kind of things, I use agents to help me to write those queries and to create charts and all those kind of things. So I'm not, you know, writing SQL queries myself anymore. I'm asking Copilot to do that for me, and I can do that in the CLI or in Versus Code, and it's just an amazing experience. So I really push you, like, if you're just using agents for coding, think about how else you can use them as well. Another tip is experiment with models. Like, when new models come out, make sure you try them. They're getting better and better all the time and, you know, you can do more and more. And you might well find that one model is great for one task or one project, and another model is great for another thing. So worth experimenting to try different things. And one last tip is, as agents are helping us to move faster, there are new bottlenecks in the software development flow. So make sure you're ready for that and be aware of it. I'm seeing, for example, that the new bottleneck becomes code review. So if we're making a lot more PRs because AI helps us to move faster, then we need stuff to help us to review faster as well. Copilot code review can help us there by giving us really quick feedback that doesn't rely on a human. But we're seeing in our teams that we're having to dedicate more time to code review, like having a dedicated person working full time as a reviewer to help stuff to move through the system quickly. One last thing before we jump into a few minutes of q and a. What am I most excited about now for the future that's coming from GitHub Copilot? The first thing that I'm really excited about is continuous AI automations. You might have seen a project that we launched a week or two ago called GitHub Agenetic Workflows. And what that's about is kind of just like you can write GitHub actions that run when you create an issue or open a pull request and all those kind of things, you'll be able to write AI automations that run on GitHub when things happen. So imagine, like, having an agent that runs once a week and updates your documentation or having an issue triage agent that when an issue comes in, gives it labels and maybe even tries to fix it. I'm super, super excited about this because it's an example of AI becoming even more autonomous and being able to do even more for us. Another thing that I'm really excited about is leveling up our mobile experience with agents. You already saw in the in a little video that I showed you, like, start telling Copilot to work on something in the cloud from my phone. We're gonna be making that even better, though. So allowing you to see the live logs from Copilot while it's working from your phone as well, and allowing you to, you know, talk to Copilot while it's working in Spirit so you can even when you're not on your computer, still have a full agent experience wherever you are. And the third thing that I'm really excited about is even more ways to hand over boring work to agents on github.com. I already showed you that you can at mention Copilot on a pull request and ask it to make changes, you know, ask it to fix failing tests. I said that that's gonna be able to push to the same PR, which I'm really excited about. There's gonna be more cool things like that. So imagine if you could have an agent that watches for merge conflicts on your PRs and automatically fixes them, so you never have to fix a merge conflict again. That's something that I'm definitely, definitely super excited about, and it's gonna be coming soon from GitHub. So that's everything that I have to share today. Thank you so much for taking the time to be here, and I hope this has been fun and interesting for you. Maybe you've seen a lot of these things already, but I hope you learned at least one thing or got inspired with one feature or one tip. We're now gonna jump into some time at q and a for the last couple of for the last couple of minutes. So excited to see what you have to ask and really happy to chat. Thank you so much. Thank you so much, Tim. We have a lot of amazing questions from the the curling and hockey fans, in the audience. Also, just wanna say that a a lot of sentiment was shared that everyone appreciate you doing a live AI demo because everyone was weighted with, nervousness demos how that would go. scary. So Live demos have always been very hard, and with AI, it's even harder because there's much more things that can go wrong. But I always love to do it live. It's much more fun, and I want to show all of you guys, like, I believe in these products. I this is how I'm using them. I don't wanna, you know, give you just videos. It's much more fun to see the real thing and and have fun together. Awesome. So let's get into some of these amazing questions. So the first one is that AgenTic CLIs have been getting a ton of investment and attention lately. Could you walk us through why a developer might reach for a CLI versus simply doing it in the IDE? Yeah. I think that's a great question. I think maybe sometimes people assume, like, when I when I talk to developers a lot, like, people people people maybe worry or think that, like, the IDE is being replaced and it's now all about the CLI. I don't really see it that way. Like, I think of it more as just two different options, and some people are gonna like one and some people are gonna like the other. You know? If you look back for the past twenty years, we've had lots of different IDEs, and we support Copilot in all of them. It's not that, you know, necessarily the s code is better than Visual Studio or better than JetBrains. It's just like we're all different. We're all humans, and we like different tools. So I think what I would say is, like, CLI and and an IDE are just kind of different ways of experimenting with agents. And depending on how you like to work and your personal style and your personal flow, you might like one or the other better. So I would suggest that you just try both and see how you get on them. We're not forcing you to move to one. We're still investing super heavily in both. You know? We have big teams working on Versus Code, big teams working on the Copilot CLI. All these things are kind of moving forward in parallel. I personally love the CLI though. Like, for me, that I think it's I think the thing that changes it most for me is that it just changes my relationship with the code. I can focus more on prompting the agent and validating what it's doing and focus less maybe on the code that it's writing at the time because I can fix that later. So it helps me to deliver things faster, and I I love that. Nice. Thank you for that. There were a lot of questions around context windows and context size. I think some other context windows of other products were cited as well as some questions around context problems with large repos or context across multiple repos. Could you talk us through how we approach the challenges of context windows with Copilot? Yeah. That's definitely a great question. Context windows are are always always a challenging thing, like running running these AI products is is expensive. And we have to, you know, like everyone, kind of figure out how we can how we can manage context windows to be able to offer the products to as many people as possible. Right now, it is true that there are some places where our context windows for some of the models are are not as big as other places. What I can say is we are already working on this area and working very actively to expand those. I can't give you a timeline or promise exactly what's gonna happen and when, but the feedback is heard really loud and really clear. And we know that people want that, and we're working on it. So you can expect in the future to have have larger context windows. On the question of large repos, I think that's also really, really important. I think the agents have got a lot better at dealing with that over time. You know, we've got really smart, like, compaction and summarization and things like that to, like, fit more into context windows. We've got, like, sub agents in Versus Code and the CLI and in the cloud so that, you know, when the agents like doing some research on the code, it can do that in an isolated context window and not, you know, vomit a million tokens into the main into your main window. So we're trying really, really hard to allow you to work on big, real projects using agents and not running to context window limitations. And you can trust that we're gonna do that because we need that for ourselves. We have big monolith code bases at GitHub. You know? GitHub's been around for, like, eighteen years now or something like that. We have code bases that have been around for all that time and then millions and millions of lines of code. So we're working really hard. Already, we're using it, these things are getting better and better. They're continuing to push on that as well. Awesome. Thank you. Taking that to the past the creating code space, how do you find the right balance between generating code and reviewing it? I think there were some questions around spectrum and development. There were some questions around, sharing the workload across team members. You talked a little bit of at the end about Copilot code review as a way to to assist with that process, but how do you think about the changing demands of shipping good code, and and where to spread that? Yeah. I think this is one of the hard things that we're seeing. You know, if I look at teams like the Copilot CLI team or the Versus Code team, these are the teams at GitHub that are the most successful in using AI and are using AI most heavily. And they're seeing they're making a lot more pull requests per week or per month than they were in the past, and that's creating a heavier load of review. I think things like Copilot code review can help there, but we actually I think we do have to change the way that we allocate our time as teams as well. We need to expect that we're gonna gonna have to spend a bit more time on code review than we did before just because there's, like, a a higher throughput in the system. Something we've been doing in some of our teams is just having a person working full time on code review, you know, Having a rotation where each day, one person's job is just to focus on code review. That's really great because it allows us to ship you know? It's it's no good coding faster if we can't ship faster. Like, we need to do both of those things at the same time, and that can help us to that can help us to do that by kind of dedicating time there. And, of course, you know, alongside Copilot code review or something like that, you can also use things like the Copilot CLI or Versus Code to help you to review as well. So asking questions about the code, helping you to understand it, that can be really useful as well for reviewing code that you don't you haven't seen before, maybe you don't understand that well. Amazing. I'd love to do one more question. Thank you for being so generous with your time. This question gets a a bit to the change in user experience from working in the IDE versus the CLI versus a cloud agent, and it's kind of a question about steering agents. So when you were working in the IDE, it was very short feedback loop of, oh, it's gone on a tangent or it's missed some piece of context or it's doing something I don't want it to do, and I'm gonna just stop it and then quickly address it. Now that we're doing things like sub agents or delegating to cloud agents and having more long running independent agents, how do you think about steering those agents or preventing the need to have to steer them in the first place? How do you keep them on the on the rails? Yeah. I think things like planning upfront are really, really important for that. Like, the more that's that's the beauty of, like, spec driven development as an idea or just using planning with agents first, which is one of the tips that I gave is that it helps to keep the agent on the rails. Like, if you have a clear description of what you want, it's much more likely you're gonna get the right thing at the end. So that's, I think, one of the biggest things for me of, like, how I how I make that work. One thing that we have been working on across all these products, you know, CLI, IDE, and cloud is having good ways for you to, like, talk to the agent while it's working as well. That wasn't in Versus Code until quite recently, but now you can interrupt. You can, like, say, oh, by the way, do this. And that's really important because we wanna make sure that if you don't have to be watching Copilot all the time. You can tab away. It can be running in the cloud and doing something in the background. But if you are watching, we wanna make sure that you can jump in and provide corrections if you want to. I think the other thing is just models are getting better and better as well, which is which is exciting. Like, we've seen huge improvements over the past six months in the in the last year just because models have got better and they get better at following instructions, better at making good decisions. And I'm sure, you know, in three months' we're gonna have even better models that take that forward further. Yes. That's that is exciting. I do think that the advancement of the models has empowered these CLIs to be even more useful than they were before as you're kind of delegating more and more. Well, cool. Thank you again, Tim. Really appreciate you walking us through all things agents. Thank you everyone for for taking the time today. It's been, yeah, really lovely, and I appreciate you, asking all these thoughtful questions. We will have a recording sent out to everyone who's registered for the session. The VOD will be available as well. So the link to the keeping up with copilot series is evergreen, and people can register in the future and watch the session as well. They didn't need to attend today. And there will be future keeping up with copilots each month, so be on the lookout for those as well. Thank you again. Thank you, Tim. Thank you, everyone. Yeah. Thank you all for being here. Really appreciate your time, and have fun with Copilot. I certainly am. Thanks so much.