GitHub Constellation India 2026 — Keynote: AI. Developers. The Future.

Heat. Heat. Heat. Please welcome J Park. Good morning. How is everyone? Let's try that again. I know it's Saturday morning. How is everyone? There we go. All right. Well, it is great to be here. This is our event, Constellation. We were here two years ago and boy, the world has changed a lot since two years ago. Now, I try to make it out here probably once a year or so to visit our teams here. And it's great to connect with them, but one of the things that's actually been really, really exciting in the last two years is actually spending time with you all with this community. And it's been great because there's so much happening and there's so much innovation here and there's also just a lot of really cool things being built here. And so you know we've seen just this really really great set of energy here in India as we spend time with the community startups larger enterprises the investors our team here GitHub and Microsoft as well but I think it's really good to look at the numbers because this really puts it in perspective so we did some data digging for this event. India is our largest developer community in GitHub with 27 million developers. Now what's really impressive is that 2 million developers have signed up for GitHub just this year. Two million. Now I really like the next slide because when you look at the contributions in open source the developers in India are the largest contributors to the open-source projects on GitHub. And now if you look at the contributions to AI projects there's been 7 and a half million contributions to AI specific projects in GitHub. Now, not only is India the largest contributor to open source projects in the world, but India is the second largest in AI contributions only to the US. Now, I don't gamble, but I'm willing to bet that this time next year that may change, right? So the other thing that I wanted to talk about is what I think is pretty special about being involved with this community and the work that we do to support all of you and lots of people in this tech ecosystem. First of all, one is it's a very high energy set of entrepreneurs in this industry. Right? when I spend time here last year and then we've done a bunch of events where the investors as well as 50 60 founders came to the Bay Area to hang out with us for two days to go deep on tech on culture on product as well. The innovation that's happening here, the entrepreneurial energy is just amazing. It's really really fun to learn from this team. The other thing that I think is also really special about India in this tech ecosystem is that the government is actually doing a lot of initiatives to transfer to transform things like healthcare, social welfare, digital infrastructure, education and lots more. Thirdly, the scale of what it takes to build a technology company here in India. not only to serve India but the aspirations from many of you to take the scale to take the business to take the tech and to scale it all around the world is also very unique and then the last part of this is that the demographic is a younger set of builders in this country and that energy and that cohort or this group of folks they're growing up they're building with AI first right so we just see in a very very large set of builders that don't know how to do it the other way. You're building with these tools and you're you are an AI kind of first developer or builder. Okay. Now, we're going to continue to support this community. We're going to work together in the open. We're going to learn together. We're going to continue to build these products, these platforms together, these open source systems that we need to do to keep up and to keep pushing the advancement in this AI era. Now I want to also kind of go back because two years ago when we had constellation we've actually been transforming or changing the way we think about and we use AI. So then two years ago I'll call it we were mostly building things or we were using AI in a way where it was question answer go back to doing work the normal way come back tomorrow question answer maybe we turn this into a chatbot we put it on our website somewhere it mostly didn't work and then we started to get improvements in the models we started to understand how to run them for maybe a little bit more sophisticated things than just a question answer. And we were able to and now most people are in this phase where they can assign a task, not a simple one, maybe a kind of a medium level task to an AI to an agent and that thing can maybe go do some research, maybe it can take some documents, summarize them, give you a proposal back, research some things on a website, maybe but these are generally you're assigning a task. You're giving it a maybe not a simple project but like a mediumsiz project usually within minutes right now which is kind of like nowish but thenish is like we're starting to see people get more and more comfortable with using agents for longer running tasks and these tasks can last many minutes or many hours or even many days in some cases. Right? So this is what's happened just in two years. We were starting off in chat kind of chat bots simple. Then we had this simple give a task to an agent and now we're actually using agents more and more autonomously in our day-to-day. Now this is still early right now. What's pulling this next phase faster and faster for us? It's the advancements in the model. It's the advancements in the tech ecosystem. And this is a chart from Meter. And you know, listen, there's lots of different horizon kind of metrics out there. This one isn't perfect, but it is pretty interesting to start to see here what's just happened in the last four to six months of what a particular set of models can do where they can run longunning tasks for many, many, many hours that replicate what a human can do with fairly high accuracy. Now, I don't think we're done here. This is going to continue to hill climb, right? In a year from now, what we're seeing here is going to look almost flat to us. Now, the other thing that we see is this stuff is not just theoretical. People are actually using this. So, here's the impact back on the GitHub platform. And this is why the GitHub team is so hard at work dayto-day. We have seen just a massive explosion in growth in GitHub. So, I want to go through here because commits Last year we did a total of about a billion commits, right? The entire year, 365 days, a billion commits. Not too shabby, right? Couple weeks ago, we did 275 million commits in a single week. So that is just like an explosion of scale. Right now, the other thing that I think is really impressive is GitHub actions. You guys all use that for CI/CD for your kind of builds and whatever you're doing to get your applications out into production. Couple years ago, we had about 500 million minutes in a given week of GitHub actions. Last year, there's about a billion per week. Last week, we hit 2.2 2 billion minutes in a single week. So again, just like massive explosion. Why is this happening? It's because folks are adopting AI and using this as their daily driver to fix things, to build new things, to experiment with things, to transform their businesses. Right? Okay. So, we're going to continue to see this because this growth because we're pulling in more and more value and understanding how to use these models. These models are getting bigger. They're getting more capable. They understand how to use more tools. They know how to plug into more ecosystems. You can use skills to actually do these things. There's just a lot of choice out there. Right now, who gets to benefit from all of this, right? Because it's actually folks like us. You guys, it's builders. You're on the forefront of this. You're the types of people who will wake up on a Saturday morning, have an idea, start, you know, building this thing, iterating on it, maybe share it with your friends. Maybe you come and ruin your Saturday morning to come here and listen to me drone on. And you know, you're you care about your tools. You care about the the craft of what you build. You're picky about which tools you use, right? And you care about your workflows and you're constantly trying to make it better. You're trying to think, what are better ways that I can unlock my imagination, my creativity, my individual creativity, or my collaboration with my teammates, whether it be in school, whether it be in work, or whether it just be some friends of yours. Now, the one thing that I don't think we talk enough about though is in the last two years, we've been riding on these three assumptions. And these three assumptions, and I'll walk through them one by one, are largely disappearing. They're all cracking. They're all breaking. They're no longer true. Now, the first assumption is that we always thought that, hey, you can take AI, you can put it in one surface that we use and you kind of just are stuck using it there. We can put it in a chat pane in an IDE. I can put it on some website. Maybe I put it in the terminal or I put it in another productivity application. But that's not how we work, right? We're at work. We talk to people. Sometimes we're at our desk. We're doing individual work. We at home maybe doing some work, side projects. Maybe we want to work when we're on the train commuting to work. Right now, you can't use AI today as you move throughout those surfaces. Right? So, this assumption that AI is only in one pane and one surface is breaking. The second assumption here is that you had to pick a model and you had to use that model forever for your whole task. Not maybe forever, but for a long time. And then you couldn't take advantage of all of the different types of models out there. These models are getting better, but they're not getting better in just one direction. Some models are small. They might be cheap. They might be fast. Others might be big, slow, but very, very capable. Others may be specialized for images. Others might be specialized for voice. And so why can't you take advantage of this whole gamut, this variety of different models for the type of thing that you're trying to build based on your imagination. Now the third thing is that in the past we've always had it so that the developer, the builder is always driving. We send it a question, gives us an answer. We give the agent a task, it gives us something done. But now with these models being more and more advanced and the rails are more capable so that these agents can run for very long periods of time where I'm not actually driving the agent, the agent is doing a task. Maybe it needs to check back with me somewhere along the way or it's going to report at the end of many hours that it's done with whatever long complex task that I gave it. Now, here's the thing. We've been running on these three assumptions for the last couple of years of this is being kind of the state, but all of the assumptions are breaking. Now, what happens here is that we can't iterate. We can't patch these. We can't think that, hey, we'll just tweak these assumptions and we'll be fine. Because what's happening is as I have AI spread through more and more of my tool base in my tools and I can take it wherever I am and continue my work, right? One of my like favorite things to do is when I'm out and about, if I have an idea for one of like my projects is I just bring up the GitHub mobile app and I just give it the agent a task and then it does its thing and then when I get back to the house there's a PR for me to review, right? and then I look at it, I fix it up and then I merge it. And I can just like have ideas wherever I'm wherever I'm at and I can fire these things away. I can also be in my desktop and say, "Hey, I actually need to I'm moving somewhere. I'm like traveling. I want to put this into the cloud so that I can keep an eye on it and steer it wherever I'm traveling." So these are the types of things that we're really trying to push is like bring that that surface that AI wherever you are and then as I talked about we want to have access to all of these models and we want to be able to run these agents the more and more autonomously but what this happens is when you see these three all come together is there's a compounding system that is starting to build here right that compounding system has one fundamental point what it's going to change and that fundamental point is that our entire lives well like I don't know a few decades here we've been thinking about and we always talk about the software development life cycle the SDLC right and that entire SDLC that entire life cycle because of this compounding system collapses now there is no such thing as the SDLC in this AI powered world of builders And we can't iterate our way there, our way there and come up with a new version of the SDLC. We actually have to reimagine and rethink this entire process of how we build because what happens if any of us can be doing ten you know tens dozens of PRs a day. What happens to how you merge these things, how you test these things? What do you do about CI/CD? What about documentation? What about security, vulnerabilities, bugs? And by the way, what does even tech debt mean anymore in this world? It changes our definition and understanding of tech debt is in a backlog. What's a backlog these days, right? So these are all going to force us to really rethink this entire ecosystem. This entire process, this SDLC collapses and something else is emerging. What is emerging is the focus of our team at Microsoft, the coreai team which includes GitHub. This is our entire focus is to come up with the new system of tools and platform to enable you all to build for the future. Now we think about this across this spectrum of building. What do you need to build? Take an idea, be able to experiment, build these things, be able to deploy them and then operate them. Where do these applications, these AI applications run? They need a different platform. They need observability. They need security. They need responsible AI. They need to be able to be debugged. They need to be able to be patched quickly, etc. Right? So, with that, I wanted to highlight some of the things that you're going to see today. And we are going to kind of just start with a quick, I think, summary of the stuff that we've already offered today. So you're going to see how we tie working across the ID in Visual Studio Code along with the CLI and github.com. So you're going to see how you can pull these three things together today. We have some other things coming that weren't quite ready today, but you'll see them soon. And then you are, I think, some of the first people outside of the four walls of our team to see this new GitHub application that we're going to show you today. So, we haven't shown this to anybody else. I hope you're excited. We still have a lot of work to do on it, but I'm really stoked about this thing. And the other thing that's really neat about this is it actually takes the same agent loop in the GitHub CLI SDK. So think about that agent loop that's actually embedded in the overall AD the agentic developer experience here. So those things come together and then we need to have a platform where we take these applications that we're building. We're not building any more kind of static websites, right? We're building applications that use agents and we want those entire applications the agents with them to have the modern the platform for those agents to function correctly. So we're going to talk about Foundry and what Foundry can do with things like toolbox, what it can do in terms of self-improvement, what it can do to hook into your knowledge bases at work. Think about your documents, your emails, etc. like that. so that we can make these agents more and more useful. Okay, so enough of me talking. We're going to show you some stuff. I hope you're excited. And with that, I'm going to welcome Karan on the stage to get us started. Thank you, J. Thank you. All right. really great to be here at another constellation sharing more about you know what's coming up and what is the thing that all of you are using as well. All right so one thing that I can tell you is that there's so much happening and all of those tools be the ID or the CLI or even github.com that's happening with copilot. All right so let me show you what we can do with it and what's shipping new. Okay, so we have our uh favorite VS code out over here and I'm on VS Code insiders so I can show you the latest and the greatest as well. And uh you know I thought that well I really wanted to build a kind of a really small web game. I'm not a game developer but I think it's something which is going to be interesting. So uh what I did is uh you know I built a small web game which I like to call it as the roadrunner. All right. So an infinite runner. Now the first thing that you might notice is that I now have a really good browsing experience right here within VS Code. So I can just you know kind of do whatever I want over here and then uh you know play this game and uh you know use my browser right over here as well and kind of do a bunch of things. All right. Now uh say I want to add something more. I want to add you know a few more things on how I can make this better. Um, you know, say for example, now the character is moving around, but I want to kind have like a small motion trail effect or something of that sort. And of course, I'm going to use copilot's help for it. So, I will bring up copilot over here. And uh, you know, you can of course see the familiar uh, panel over here with so much more that you can do. you have a choice of so many different models uh you know be it uh you know many of the cloud family models, open AAI models uh you know Gemini models etc and you can even use and bring your own models uh there are various different agents um and various different modes that you can use um and also you know you can use it in uh locally on the CLI or the cloud as well. So I'm going to just uh you know go with uh uh local right now. I'll keep it opus 4.6 medium and I will ask it to say that you know add this brief color motion trail etc. Now one of the things that's really really useful while working in the browser as well as copilot is I can now actually share my browser right here with the agent so that copilot can drive the browser for the complete build and test and validate cycle as well. So I'm going to share that with the agent. And another really interesting thing which is out here in preview is the permissions. So if I go and see over here there's default approvals. You can bypass all approvals uh and an autopilot mode. So it itates right from the starting to the end on its own. So I'm just going to get into the autopilot mode and uh I'm going to give it an instruction to test it out in the browser and have copilot work on it. So, it's going to get, you know, uh, started trying to figure out what needs to be done, just like how Copilot works. Now, while it's making the changes, I want to show you a few more really powerful customizations that you can do with Copilot as well. So, if you go right over here, you can see chat customizations. All right. So, what this includes is you're able to customize agents. What are the agents which are available uh for you? You can generate a new agent. Uh you can even use skills. So you can see I have a play game skill which actually tells Copilot how to play this game by itself. Uh custom instructions, prompts, hooks, MCP servers, plugins and other things as well which is really useful when you want to give Copilot the right set of context, right set of skills and you know be able to do how uh you can use it. Now, so you can see that I'm not doing this. All right, this is copilot playing. Oops. So, it hit but it is testing. All right, so game is playing. Uh, and you can see that copilot is actually driving the browser here within the agent loop itself. All right, so it's figuring out the changes, paused it right over there. Uh, and seeing that, okay, it's taken a screenshot and figuring out that all right, the trail effect is working and a colorful motion. All right. Now if you think about how you would have done this all right uh is probably in a very separate manner but you can bring this together right here with uh VS code itself where you can have your integrated browser and you can see copilot make those changes drive the browser test and validate that all of those changes work fine and then you know complete the task that's amazing right of what copilot can do with the integrated browser in VS Code All right, let's keep that. And uh you know, one of the other things that uh I really want to do is accomplish a whole lot more tasks. Now, last year at Universe, we announced agent HQ where you can use multiple different agents from different providers right on github.com with copilot. So what that means is I can go right over here and switch to a cloud agent and I can use cloud agent or codeex agent with copilot itself. So I can say switch to cloud agent and I can ask it you know for say one other uh chain saying that add some idle animations etc. Now there are some changes I'm not going to commit that but I will just delegate it straight away. So what this will do is it will take all of the context and hand it over to copilot which will run the cloud agent on the cloud and create a pull request for you when it's ready. Not just that of course you know I can even go back and then switch to uh say codeex agent ask it for uh you know another change that I want say uh you know add like a real movie style dust cloud when the player lands you know bof kind of a thing and make all of those changes whatever u you know whatever I need whatever I want as well. So uh I can see that all of these sessions are running right here in the sessions panel where you can see this that's running on the cloud what we completed right now and many other sessions as well. Not just that I can even you know delegate some of this to copilot CLI to run in the background. So uh I can just say that um I'm going to select copilot CLI and also the same sort of models and agents. Uh but now I can also of course mention whether I want this to happen in the workspace itself or isolate it in a work tree. So I will ask it that well what's a roadrunner game without some powerups right coins are boring. Um, I think what would be really interesting is if those collectible were samosas. No, go over there, collect some samosas, get some energy, and go on. So, I'll just ask for that. And again, I'll have it run on autopilot. Um, since it's in a work tree, I'll copy over the changes and let it go. Now, of course, you know, you can use copilot CLI right over here in the background or I can even just open up a terminal and you know start copilot right from here as well. All right. Now, one of the amazing things is that of course copilot is here for you wherever you are be it in the CLI, VS code or github.com which means the session that I kicked off out over here I can actually see those sessions. as you can see over here the collectible samosas which is in use right in the CLI as well. So it really helps you bring together all of those tasks wherever you started um you know and you can continue that with copilot itself. Now I know this is uh you know a bit of a small screen for the terminal. So let's go back to uh you know a good classic terminal where I can open up my item and uh continue with copilot cli. So, I'm going to get into uh copilot over here in the CLI. And just because, you know, I like to work a bit dangerously, I'm going to turn on yolo mode uh which basically provides all permissions to do whatever you want. Uh now since you know if many of you might have used uh copilot uh CLI at an older version or so there's so many changes you know that we are shipping and if you just want a you know a really good highlight of what's new you can just use the change lock command and you know ask it to summarize something like you know summarize since 0.4.0 Oh, and it's going to look at all of those change logs and give you a really nice summary that well there's a whole lot of new stuff uh which you know you can do a bunch of things like share HTML um and all of those things that you can you can do as well. All right. So really really uh you know cool. Now of course I can uh change my mode as well uh whether I want to run this in execution or plan or anything else. So I'm I'm just going to u you know switch to the plan mode right now because I want you know to implement one other feature in the game which might require some planning. All right now um you might notice that I'm adding many more features in the games. If there's anything that comes to your mind, keep it in mind. All right. Probably something that we can build on as well. So I'm going to go ahead and then you know ask sketch to do an India themed powerup system. All right where players can go and then get one one cutting chai one filter coffee one jugard shield etc. So the game flows you know really well. So that is going to go and then figure out the whole code base what needs to be done uh what is already existing etc. And while that's running, I want to show you a few more really interesting things in the uh in the CLI. So I will of course allow this because it's uh you know still exploring with CLI on VS Code. So I will just um I will just go ahead and bring up uh okay yeah it's trying to find a whole lot of different files not here elsewhere. So I will just bring up the question mark which shows me the shortcuts. Okay. So really nice. I can see all of the keyboard shortcuts, many of the slash commands, you know, like slash agent/skills, u many of the slash commands related to models and sub aents like model, uh delegate, fleet, task, etc. Um code, permissions, session, uh a whole lot of help and feedback as well, etc. Right? So, you know, there's there's so much that you can do and of course, you get the same model choice when you're working over here. Uh and you can just do slashmodel to know what are the models you have access to which are the ones you want to um get get a sense of of how to use. All right. So it looks like my plan is uh you know coming close to an end. So you can see that it has figured out most of the things of how to um write the powerup manager and what needs to be done. So it's uh it's coming up uh with the plan um as well. So I can also switch to various different modes like the autopilot mode right here as well. So that just like how we saw on VS code uh you know we can execute a task completely with autopilot uh as well. So uh the plan is coming up and uh you know you can see that I'm working on multiple different agents. I have two or three agents running in VS Code on the cloud. I had a CLI agent working in VS Code. Um, and of course I'm right now in the CLI as well. And you can see that it's also created multiple of these todos of how to go about uh implementing uh the plan and what are the specific todos that it needs to do and uh you know it gives me the the plan and then what should I be doing accept it autopilot default permissions or should I prompt myself? Well, autopilot would be really good but I want to show you something else. So I will just go and then say I'll prompt it myself. All right. Now an interesting thing is that you see there were like 17 todos u but that were over there to implement right with the new fleet mode in copilot CLI. What I can do is I can have parallel sub aents go and execute what I want. So what I can do is I can just go and then say implement the plan within a new work tree and give it into the fleet mode. So it goes and then starts looking at todos uh what are the dependencies, what needs to be accomplished first, what needs to happen uh later and then kind of figures out uh you know what's what's that order like and how to dispatch those sub aents. So uh you can see that you know the work tree already exists um and you know it's going at the implementation looking at what are the changes what needs to be done etc and it will go and then uh you know deploy multiple agents so uh you know it it implementation looks complete because let's be honest I was so excited I was trying to do this beforehand and then you know show you that all the nice stuff that copilot uh did with the game as well, but that's fine. It's going to go and then deploy a few uh agents as well and then I trade uh it trade through uh that as well. All right, so it's great. Everything is good and uh you know it's going to go and then well I've saved my agents for some of the better tasks that are needed. All right. So while um you know this is going on there is uh you know one other one other thing that I want to show you which is uh you know of course I'm going to bring up my uh copilage over here and um I will again go back into my yellow mode. So I want to do a few things. I'm going to switch to u shell mode so I can you know execute commands right over here. And uh you know I'm just going to do uh get checkout and you know something like my um weather system which was something that I was working on to add like a weather system to the game and I really want to kind of have copilot review this code for me. So what I can do is now use the slash review uh slash command and have it review this. So I can you know just say run this code review agent. Um or I can even say you know do this against a specific PR as well. So this I can see is on PR5. So I'm going to say review this using PR5. But now I can also say that review this using opus 4.6 and you know say for example GBT 5.4. So I can ask it to review not just using the model that I'm currently using but I can say that well you know go ahead and do a review of multiple uh things and it will say it will launch two different uh code review um you know agents as well. Um I think I probably by mistake closed uh PR number five. Uh but that's that's fine. you know, it's going to go and then launch the code reviews uh for the specific branch because I'm already there uh on that branch and it will have these two agents dispatch 4.6 5.4 for get whatever is needed from you know these two agents and once it's done it will come back and then share all of the findings right so it's not just about review from one model but across the models and and uh you know what you can what you can do with it um as well so really really uh interesting and uh nice now while this uh review is going on there is you know one other thing that came to my mind which is well you would have seen the the game it was kind of very basic right there was uh there wasn't really any realism in terms of how the overall scene looked like how the player looked like etc like I said I'm not a game developer but I'm really interested in knowing how that might work and what are the things I need to keep in mind so for that there is a slash command called slash research which will actually dive a whole lot more deeper into everything that can be done with uh copilot and whatever is your task. So um for example I can ask it not to implement but I can ask it you know to do something like this one. How do open source browser 3D games really make their uh you know procedural worlds feel alive? Now many of the times when you're working with chat right it's optimized of course for uh faster responses but sometimes you want to dive a whole lot more deeper as well into certain aspects like how it does. So the slash research command doesn't just look at your codebase but it actually goes looks at uh the web you know other uh uh projects on GitHub your own internal projects in the organization um and then you know creates a very comprehensive research on the topic that you have created. Now this of course is going to take a bit of time. So I did it uh beforehand to show you. Uh and once that's done you can actually ask it to share this as well as a file HTML or a gist um so that you can take a look at that research again as well. So I exported this research as a gist u you know a bit earlier. So I'm going to go and then show you that just right over here. All right. So you can see uh you know it gives me a table of contents what the research says matters for immersion uh you know what does it how does it matter for low poly games uh and you know it gives me uh a few references as well common patterns across a few different games about how this works um and also because I asked it for gap it does a gap analysis there is no audio and things like that um and then you know it tells me also about uh recommendations in terms of priority. This is the most tier one. There's no audio. So, kind of do that. And uh you know, if I go all the way down again, there's all of these recommendations. I'm scrolling and I'm scrolling. And that's how detailed the research is. It gives me an architecture diagram as well with proposed additions as well. confidence assessment that all right you know this feels really high confidence for me and you know just like a good researcher it also gives you citations about everything with footnotes all right isn't that so powerful slash research with just that one prompt goes and explores all of those things and helps you export that artifact in a file or HTML or a gist etc. So you can actually go back and then um you know look at it, work on it and do a whole lot more as well. Right. Pretty awesome. Now uh since we are on github.com, let's kind of go back on GitHub and see what are the things that we can do on github.com. So this is my private repo and if I go into pull requests you can see multiple pull requests that I had started off with cloud and um you know codeex as well uh etc which uh is probably running or completed. Well yes it's completed it has asked for my review etc. And uh you know if I go back into you know view session for any of the sessions it takes me to this new agents tab right over here in the repository where all of the agents are uh working and you know you can take a look at what's happening. This is um you know the cloud agent which was working as well and there's so much happening um you know there as well. So, uh, the agents tab really gives you an, uh, you know, a mission control kind of an overview for that project as to all of the agents that are going on. And not just that, uh, you can also configure the copilot cloud agents to do a whole lot more. For example, if I go into configure, I can u, you know, enable firewall. uh I can ensure that you know whenever there is any code that is being pushed that should uh require approval for workflow runs and also configure validation tools. So it's not just that the cloud agent went and then wrote something but it's actually running all of these validation tools like code scanning, code review, secret scanning and everything else so that uh you know it's very much confident about the code it's writing and also have uh the MCP servers that you want to configure as well etc. So pretty cool right now. Of course, in the past few minutes, you know, we we spun off so many different agents to work on so many different tasks. It's going to take a bit of time. So, I did all of this beforehand as well and then put it all together in one branch. So, let's see how that would look like when all of these agents would have completed. So, I will go back again into my um VS Code Insiders over here and I will bring up uh my terminal again and um I'm just going to stop this server. I will check out into the branch that I need uh which is feed/final. Um and uh you know local changes that's fine. I'm just going to let go of all of the local changes. Well, I think I should do a better job at writing git ignores. Um, and I'm going to go back and check out that branch again. It's running or right create and uh, and I'm just going to since there's a static one, I'll just do an npx serve np and then 333 my port number. And let's bring up that game. I'm going to close this, close my terminal, refresh this, and uh see whether some of the changes that we had are working over here. All right. So, I'm going to press space. There it is. And uh let's see. That cloud looks very uh you know, kind of a hero style, I would say. Where are my samosas? Oh, there it is. Okay, there I got my samosa. And uh you know probably I think there should be some more powerups or so as well. Anyone who has played these infinite runners know that you can kind of go on and on and on forever. But the point is not that. The point is that you know really in the past few minutes what we did was we took this small scrappy uh game/ application that we were building did a whole lot of uh dispatch many agents in VS Code took them uh into CLI did a whole lot more planning for new features um also you know deployed a whole lot more agents for implementing them reviewed a bunch of code came back into GitHub as well to do a whole lot more of those reviews as well, etc. Right? So, so much that's possible with Copilot across VS Code, across CLI, across github.com as well, but there is a whole lot more to come. And to show you more of that, I'd like to welcome Kyle Dagel. Thank you. Hello. How's everybody doing? Oh, come on. How's everybody doing? There we go. That's better. So good to be back here. I was last in India last April. And when I'm chatting with folks and chatting with so many of you, I mean, so much has changed in such a very short period of time. Just a year ago, like Jay was saying, we were really just starting to talk about Agentic AI, right? We were asking a question, waiting for it to respond. I was joking with everyone that we were all just like staring at this thing working and we were saying how productive we were while we waited for it to work. Now, these agents are working like asynchronous teammates. We're able to send them off on a task and I can run multiple different projects and pull requests all at the same time. We've been on a journey from starting like myself as a dev just writing code directly to you know working with an agent that was writing some code in a single project. But I think as we look to the future and some of the things I'm I'm going to show you, we're really talking about one person working with a bunch of agents working on a team of human beings, agents, developers working together side by side. So we're not just introducing the idea of agentic AI anymore. We're at the point where we're really scaling it. But that introduces a new kind of challenge for us, right? It's it's a good kind of challenge. We're moving so fast that we really need better tools to keep up with this new problem. So, think about what Karan just showed you, right? Multiple agents, VS Code, the CLI, github.com. But that's cognitive overload, right? How many projects can you possibly keep in your head at a single time? You're jumping between Windows, terminals, the web browser. We think that there's capability there, right? There's so much we can do, but we have an opportunity to create a better home for it. And so we asked ourselves, what does GitHub look like when it's built from scratch in this new reality? When agents aren't really a feature you opt into, but a regular part of your day-to-day work. And so the team at GitHub knows one of my favorite things is to show things that are works in progress. Jay told you no one outside of GitHub has seen this yet. We're bringing a special preview to you. Now, to be fair, we're not quite ready to give this to all of you. So, you're developers, we're developers. We're going to show you a work in progress, and that's this new GitHub experience in an agent first world. This is workflow aare agents together. It assumes that you're going to be working on multiple projects at a single time. It brings together a Gentic development and the GitHub platform that you already know and love all into a single place. So, it wouldn't quite be fair if I didn't show you a live demo. So, let's give it a shot. Okay. So, this is the app. I think I can't see the screen. You can see the app. Yeah. Yeah. Great. There we go. Hey, hey, there we go. That's my favorite. I'll start here. Don't worry, she'll come back. I promise. So, this is the GitHub app. As you can see, it looks very similar to, you know, a prompt that you would start with. I can choose the modes like Karan said. I can choose the models. I can choose thinking uh the thinking level that I want the project. I can start with work trees, choose the branch. I can go so far as to add from GitHub a local repository or create a new repo all together. I can also choose what skills I want to use, MCP servers, so on and so forth. Now because this is also a part of you know your daytoday workflows I also have all of my pull requests. So this is a pull request you know to implement double jump inside the roadrunner uh app that we were working on. I can click into a few more that we were looking at and at any point I can just come up here and click create session to start working on this particular pull request. Of course, I also have all of my issues. Very fast to go between all of these. And you know, last but not least, of course, I can just chat, right? We always are kind of working on something and want to ask a quick question. And so I could just come in here and say, okay, what's happening at GitHub constellation right now? Now, this agent under the hood is ultimately using the same harness and SDK as the C-pilot CLI and so much of what else we're building. And so while that cooks for a minute, I want to show you my favorite feature when I'm using this uh every single day. And that is workflows. So workflows gives me the opportunity to tell the uh the app to run a particular prompt uh on a particular schedule. So I can come in here and like for personally I do a daily briefing every morning where I go through and it looks through all my GitHub issues, works through my email, goes through everything. And I could just create a prompt here for that. Select the projects that I want in the models. Or I could just go ahead and you know use the existing workflow that was created. And this is an issue triage workflow where if anyone works in open source or in a larger company, right? I feel like you wake up, you get your coffee or your tea, and then you read through all your notifications. Uh and so in this way I can just get started with exactly what I want to start my day with which is usually you know either reviewing or getting right into writing the code. And so now if I come over to these other projects I can uh take a look at you know projects that were I'm sorry pull requests that are already running for that roadrunner app or I can go ahead and create a brand new session which will create behind the scenes a new work tree for me so I could run multiple at the same time. And I'm gonna add a screen counter that shows uh meters and kilometers run instead of uh uh just sort of letting it not count up and just show you the various pieces that like that Karan showed catching all of the samosas and whatnot. But so I don't waste too much time because I know we're running a little long. Uh I want to show you an example of a pull request that was already completed. So you can see in here when we sent the app through it went through and looked at all the code. Like I said, it's thinking, it's making the implementation details. And then on this side, I can actually take a look and review the code right here. I can look at the pull request in case there's been any changes. I can also just look at the plan that was written by the agent in case I wanted to make changes. I can also, you know, open this in VS Code or other editors. And last but not least, you know, why wouldn't I just start running the actual application myself? So I can start the application right in here. I can open up my browser. We'll jump to it. Local host 1 2 3 4. And then I can just play the double jump PR that happened. So if I get going Yep. And the double jump works all within this single environment. And now when I want to go ahead and uh merge this, I no longer have to sort of jump out. I can click merge here or I can go ahead and actually enable agent merge which will handle the merge for me, deal with any conflicts, resolve any comments that came in from co-workers or from other agents and quickly get me back to work. So that's a very very quick demo of this new GitHub experience that we're building. What did you think? Yeah. And so if we can go back to the slides, like I was saying, I mean, the app that we built here is still early. We're still very much cooking, but we're using the app to build the app and ship the app. And so we're going to have a lot more to share with all of you very, very soon. But was very excited to give you all the world's first sneak peek at what we're cooking up over here. Like I mentioned that app and so many of our experiences run on this singular agent loop, right? It works the same way. Uh and it's all built on top of our C-pilot SDK. So whether you're using the Copilot CLI or VS Code, github.com and our cloud agents or this new GitHub experience, it's all backed by that one single thing, the Copilot SDK. And a question we kept getting from developers was, you know, can I bring GitHub C-Pilot into my own applications or at work? And we've recently public previewed your ability to use the Copilot SDK in your own applications. So now you can take this SDK, have the same model choice, have access to the GHLI or MCP servers, the same session management, being able to bring sessions across the different interfaces instead of being stuck in just one. And so now C-pilot isn't just something that only lives in the GitHub experiences. It's a foundation that all of you can build on as well. So I want to show you a real quick uh demo. I think I'm back. Let's see. Okay, here we go. Great. So, I figured let's go back. And we created a really quick app here that basically just emulates, you know, uh, Windows XP. And so I quickly can come in here and do some things, but we thought, wouldn't it be cool if instead of just, you know, emulating this, uh, we went ahead and made GitHub Copilot, uh, a part of this. So, I'm going to check it out. I'm going to restart the server real quick. Oops. All right. Give it a start. Never gets old. Okay, I'm going to come back in. And now as it loads up, I can go ahead and come back to that command prompt that I was looking at. And we've actually embedded Copilot inside of this prompt. So now I can just say, you know, hey, uh, what's your favorite programming language? People are watching. Uh, and you can see, you know, it's working over here. Uh, uh, it's going to take a second cuz it's in this alt interface, but yep, keep it professional. Oh, wow. That's weak. Uh, okay. Uh, but ultimately what we're see there we go. See, just got to give them a nudge. But really what we're looking at here is we're just looking at a simple, you know, React invite application that embeds the Copilot SDK into this experience. You can imagine for all of you whether that be in a side project or at work particularly if you're want to build that sort of custom co-pilot experience uh instead of having just some fun with Windows XP you can bring C-pilot there and you're able to actually then see this session like I was saying across all of your various uh interfaces as well. So this is the session I just kicked off, right? This, hey, what's your favorite programming language? Even though it was inside of this SDK experience. So those will follow you through the rest of your uh usage of Copilot, whether it be CLI, VS Code SDK. And then you can just see right here real fast. This was the prompt that ultimately created uh the experience with the SDK using autopilot. There you go. Okay. So now with the power of this SDK across all of our various interfaces, you can build with agents. You can take those agents into your own applications and you can work alongside them across any surface that you're using. But one of the things that we frequently get asked by our customers is how do you go about doing this and building and deploying agents at scale with governance, security, all the things that our employers are going to ask us to do, not just the stuff we get to do on the weekends. And so to answer that question, I'd love to turn it over to Tina. Tia. Hy. Thanks. Hi everyone. My name is Tina and I'm super excited to be here. What an amazing audience. So, you just heard from Hy how easy it is to build agents and manage agents. But what if now you need to share your agents with friends and family and co-workers and host it in the cloud so that it's always running? That's where Foundry, that's where Microsoft Foundry comes in. Microsoft Foundry is an enterprisegrade developer platform per purpose-built for building, deploying, and operating agentic applications and closing the loop by optimizing the whole cycle. We're seeing a lot of our developers taking advantage of harnesses such as GitHub copilot SDK or cloud agent SDK and that's a great start but when they need to deploy it into production they often run into a wall they end up have to rebuild the same infrastructure again and again. Let me show you what I mean. So often you're going to have the harness which runs the agent loop. It does the reasoning and planning. It has access to file system the terminal and it manages the agent state. All of the all of this it's untrusted. Usually you want to run it in a sandbox. Then you also have the trusted platform layer where you have to store the identity. You have to do the credential management and you have the egress proxy there for observability and policy enforcement. But all of this needs to be in the trusted layer. So then it can call enterprise systems and APIs. So this is the emerging pattern that we're seeing. You both have the agent reasoning loop that has to be in a untrusted layer. Then you have the trusted layer which has the identity, identity, credential management, policy and external access are all in the trusted layer. Foundry is great to solve this pattern. So let me show you how Foundry had the Foundry's engine room and foundation is called hosted agent. It offers a fully flexible infrastructure. So anyway you're building the agent hosted agents can run it for you. We offer three big value props. Number one governance. We have session isolation as well as egress control. So governance comes out of the box for you. Number two efficiency. We have stateful session storage as well as scale to zero. So when you're not using the compute you don't have to pay for it. and three performance we offer sub-second cold start so your your agent is always there for you then foundry hosted agent is built on top of microVM sandbox to offer the the safe session isolation and then we build the agent glue on top of this microVM such as the control plane the data plane observability and evals as well as identity and tools so anything that you need to make a successful agent such as the uh agent version life cycle, the session management, the distributed tracing and logs as well as the identity and on behalf of for tools all come out of the box for you and all of this sits behind the foundry service gateway. So it does the right monitoring as well as policy enforcement and the end user can interact directly with the agent endpoint or through Microsoft teams with our Microsoft teams integration. As you can see here, hosted agents really offers this fully flexible layer so that whatever you however you're building your agent, hosted agents can run that for you. With that, I'm going to invite up uh Vivec, my colleague, who's going to come up here to show you a hosted agents demo. You're gonna you're going to see how easy it is to build an agent with GitHub CLI. You're going to see how easy it is to deploy it right into Microsoft Foundry. And you're going to see how much powerful it is once you have a hosted Foundry agent. Take it away, VC. Hey, everyone. So excited to be here. I'm Vic and I'm going to try a live demo and go through build, deploy, and operate phase for an agent all under five minutes. So let's get quickly Okay, so I'm starting with a bit of a cheat code. I sort of already have my code here. You can see um I'm going to quickly see what the code looks like. So there's only really one source Python file. I'm going to quickly walk through it and let's see the number of lines of code. It's around 70 lines of code. We're going to start from the top and get a good understanding what this code is trying to do. We are using two SDKs. One copilot SDK to build our agent and one and another one an agent server SDK to actually serve the agent. It has three sections. First section is we are creating a session if it does not exist and if it does exist reusing that session so that ensures that the next turns have the right context. The second section is really interfacing with a co-pilot agent where user ask a request we send it to the co-pilot agent and then wait for the responses. Meanwhile as it emits assistant messages the handler keeps appending them into the reply and we only send the last response back to the client. So that's all it's doing with these two section sessions. We're all done. Now we have to host or serve the request. So we are creating a server. It creates an HTTP server locally. And we have a response handler. This is really important where the request and responses now adhere to an open respon responses protocol. And what it helps with is that I can use uh open AI clients to interact with my agent. I don't have to vent my own personal client. And finally, I start the server with server.run. With this, let's quickly get started. And I'm going to start the server. So, I'm using an easydi agent CLI that foundry events out, but you can even do python uh uh python main.py, and that should also work. It'll start the agent locally on localhost 808 at port. And I'm starting in foreground. So, you can see all the logs that it's printing. to invoke it. I'll just use another terminal and try to do a local invocation. So again, I'm using the same CLI ACD AI agent and invoking the agent by asking it to tell me a joke and I'm using local port. I've sent the request. It's going to interact with the server locally who is running which is running on 808 port and I'm expecting a joke. All of the interactions through an agent handles within a session boundary. So that's why it's created a session already. Okay, we have our first joke. It's a pirate joke and it's in pirate lingo. Now that agent is working um and I can see that I want to switch to deploy mode. And this is really important. Uh this is where Foundry comes in. And what we are trying to do with that is to simplify deployment to as simple as just calling it a deploy. what it will do it is it'll package my agent put into a container registry and make it servable with a right identity right tools um and with responses protocol all baked in for me while it's doing that I'm going to switch to operate mode and this is uh and I've deployed this agent a couple of times already um this is my agent it is type hosted and it's a description that I've provided it I'm just going to quickly go into that and try to interact with that agent here. Uh, this is this happens with the live demos. I'm going to try again. I'll start walking through. I have a cache result if this doesn't work again. Okay. So, how it should look like is something like this where if I say hey, it will start showing me the session. It'll start a session and will start showing me the log stream in real time. I can also access traces while it's doing it. And this is really cool because with these traces, I can see what's going inside the agent. It can show me input and outputs and it can even show what's the LLM calls it's making, what its inner loop looks like, all in place. This is already same agent that I've talked and I when I talk more about it, it appends to the same session and shows me my joke as well. One cool feature is is the stream log stream. As you interact with it, you can actually see streams in real time for your container. Uh, and you can debug it in real time. Although it says STDR, these are all streams that I'm sending you to STDRs for just easy debugging. The key important thing here is it also has files. So as you checkpoint, you can see those files and you can restart your agents from there. Uh I'll go to the locks and here I can see all my previous sessions with that agent. Now sessions could be idle or active and this is where it's important that you you're interfacing with your agent through sessions. Now a session may not be active all the time. it has done its work and to have a compute associated with it is wasting the compute. What happens is if it's idle and it's not processing it can be the compute can be taken out but the state is still checkpointed in file stores and you can see that there's some sessions which are idle I can actually click on it I can see files and you can see that all the copilot state is checkpointed and it can resume from there at any point of time so this makes sessions durable while the compute is ephemeral we can also monitor the agents operational metrics through the real-time dashboard and you can run evaluations for your agent. While this was being demoed, let's see if our agent has been deployed. And it did deploy the version 13 successfully and we have the invoke ready for this. Ignore the post- deploy failure. It is just that I don't have credentials to list all my service principles. But the agent is deployed successfully. And if you are wondering and curious why it always check tells pirate jokes, it's because I steer it to do that. And I have a skill here that actually steers the agent to narrate a pirate's joke in a pirate lingo. With this we have a build, deploy and operate phases with foundry making it one click for deploy and mostly infrastructure uh invisible for the developers. That's the demo and the tlddr for this is if you want to share pirate jokes with others deploy to foundry. Thank you. I'm going to invite Tina back to the stage to take it forward. That is Foundry hosted agents. A infrastructure that is super flexible. Any way that you're building agents can be hosted in Foundry hosted agents. So you may be building your agent with just model APIs such as anthropics messages or open AI responses. You can host your agents in hosted agents or you may be building with a an agent framework such as Microsoft agent framework or lane graph. We can host it in hosted agents. You can also build your agent with and har a harness such as GitHub copilot SDK or cloud agent SDK. We can host it in hosted agents. Or you can build clause such as open cloud. We can also host it in hosted agents. Show of hands, how many people have heard of open cloud? Okay, I see all of you. So, we have the right audience here. Uh, so our team also looked at open cloud and came up with a bunch of learnings and we want to share with you. Number one, we think these are great learnings. Number two, we also want to show how we're hardening our platform to make it really powerful for running a open cloud as well. Four learnings. Number one, when agents are more autonomous, they're more useful and more dangerous. Agents that can act are way more powerful than agents that just advice. But every single capability expands the attack surface. This means we have to offer the right session isolation. This is why hosted agent is built on top of microVM sandbox. And we have to offer the right network isolation and on behalf of identity and egress control for observability and governance. Two agent side just uh agent side remember really becomes indispensable. Persistent state across sessions is it's what makes channel it's more what makes uh sorry persistent state across sessions and channels. It's what makes an agent really is like a teammate rather than a stateless tool. Persistent session state management has to be a first class citizen. We already have conversation state management such as wppy responses API. We also have agent states through disk persistence such as snapshotting and restoring. And we already have foundry memory for preference and summarization. We're also adding procedural memory so that our agents can remember how to do things such as a business process. And three channels matter. Claws really shine on meeting users where where they are such as connecting to WhatsApp, Slack, and native apps. for Foundry. What this means is that we got to be connected to Teams, M365, and Azure apps, which which can then talk to other apps and channels. We already have three IQ integrations, Work IQ, Fabric IQ, and Foundry IQ. So, anything that you need for your work knowledge is right there in your fing at your fingertips. And we already have MCP tool interrupt, but we got to make that way easier. In the next slide, we're actually going to talk about Foundry toolbox, which makes tool calling a lot easier. And fourth, best agents are proactive just like humans. They don't wait for prompts, they wake up on events and schedules. This means we got to offer event-driven execution, scale to zero, and crown job scheduling. We're building all of those. Cool. One of the one of the challenges our agent developers face is really about calling the right tool. Making sure the agents call the right tool at the right time. This is a problem Foundry toolbox solves. Curation, search and governance. Number one, curation. The developer can browse the tool catalog, pick tools and bundle them into a toolbox. He then can configure o and steering instructions. So the toolbox can be reused across agents. Number two, search. Toolbox has a tool search functionality. So it dynamically finds the right tool in context. It solves the context window explosion problem. And three, governance. Foundry toolbox offers private tool catalog. So you can set org level permissions at the toolbox boundary. So data loss prevention, RAI guard rails as well rate limits all come out of the box for you. And of course every single tool invocation is logged. and tool authentication is done through API keys OS and user impersonation that's foundry toolbox another problem hard problem that our agent developers encounter is really about detecting agent drift and have the agent self-improve today debugging is really manual and reactive our agent developers may care about five things quality cost, latency, reliability, and determinism. And sometimes they all trade off against each other. What we really need is a systematic compounding loop with one button called optimize. And that is a foundry agent optimization service solves. It solves this problem through four simple steps. Number one, observe. Foundry observability platform already offers hotel traces. So we already log the tool calls, LM turns, latency, and token counts. The optimization service adds a feedback API so developers and end users can add in human ground truth such as thumbs up and thumbs down. Number two, evaluate. offers LM as a judge and custom scoring endpoints. Foundry optimization service adds autogenerated rubrics so developers don't have to start from scratch. And three, optimize. The optimization service takes in traces and criteria and runs a reflective loop by changing agent configuration such as the prompt, the skills, the tool configuration or the model configuration while replaying all the tasks. It then picks the win uh it then ranks the candidates based on accuracy, cost and latency. And the fourth step, once a developer picks a winning variant, it automatically deploys it with AB test. So rollback is se seamless when needed. And let me show you how easy the developer experiences. Three simple steps. Number one, add an optimizable decorator to tell us which agent you're trying to optimize for. Number two, call agent.optimize optimize to explic explicitly trigger the optimization and three call agent.promote to promote the winning variant to prod. Let me show you some early results. These are really exciting 60% improvement and this is just by changing the configurations. Right? I mentioned about four types of configurations. the agent prompt, the skills, the tool configuration as well as the model configuration. So you don't need a data science team and you don't need to do any sophisticated RL just by using the optimization service and doing configuration optimization you can have massive gains and the best part is that it does this all automatically. So let me take a step back and talk about Foundry agent service. We have this fully flexible layer which is foundry hosted agent. It's fully flexible infrastructure let you run any agent however you want to build your agent. Then we have these modular components such as the foundry toolbox or foundry optimization service. So if you're using the foundry agent service everything including the infrastructure and the modular component come out of the box for you. Or if you want to use your own compute and you just want to use these modular components alakar, you can do that as well. We embrace developer choice. Remember this emergent uh product architecture we're talking about. Remember there's the agent reasoning loop that is untrusted. Then there's a trusted platform layer that has and external access. All has to be trusted. Let me go back to how Founding can solve this problem for you. Foundry agent service spons. So the agent harness then does the reasoning, does the files access, does the terminal access and manages agent states. The agent harness is not trusted. It needs to run unsecure code. No credential gets stored there. Then we have the entra agent ID that's attached to each session, but it runs outside of the agent harness. This agent ID is then used to authenticate and authorize Foundry toolbox which access authorized tools uh MCP servers as well as knowledge bases. We also have the egress proxy that does the observability and policy enforcement. All of these are running in the trusted layer. This is how Foundry packages sandbox identity toolbox and observability all in the trusted platform developer as you just bring in the hardness and we run it for you. Let's build. Okay, now let me welcome back to the stage Jay, the one only Jay. All right, what' you all think? Yeah. Okay, so show of hands. How many folks are on GitHub these days? Pretty much everyone. Good. How many folks use VS Code? Okay, you can do a little better. How about the CLI? How many people are going to try the CLI? There we go. And how about everybody try the SDK and build your own application using the SDK? All right. Well, I hope today that we were able to just give you a really, really fast tour of how we're reimagining the way we think about development in this new phase of AI development. Now, our core focus in the core AI team is to empower every developer, but more and more that term developer is changing into the term builder. Anybody can be a builder, right? and to empower builders to shape the future with AI. But in order to do that, we need to reimagine the set of tools that you use. You need to have AI through all of the tools and the surfaces that you use. You need to be able to bring that continuity of creativity and collaborate uh collaboration. And then you need to be able to deploy this on a platform that is secure, that's scalable, that's trustworthy, that's observable if you want to actually build things that you're going to provide to businesses. Now as I said earlier what's happening with this compounding growth engine of the models of the harness of the skills of open source that this entire SDLC that we've known for many many decades is totally collapsing right so from a mindset perspective I want everybody here to think about that right because no longer is the work that we do this serial set of things that we do you saw from Karan and Kyle and Tina and Vivac that actually we can do lots of work in parallel we can do lots of different type of work in parallel. So I wanted to just say thank you all for checking out all of this new stuff. Stay tuned because this is just a small glimpse of the work that we are doing in our team and I want to continue the conversation with my friend Shakar. So we're going to talk about a bunch of different things and Shikhar is the managing partner at Excel here in India and he has a lot of perspective and thoughts about what's happening in obviously the startup community and in Indian in the Indian technology ecosystem at large. So Shakar, hey J. Hey, great to see you. Great to see you. Lots of familiar faces today. Yeah, they are they are awaiting lunch, I think. So um so listen, uh I was just reflecting actually. We've known each other for a while but last year about this time I came to India and we had a dinner a bunch of investors. Yes. And it was very lively and it was a very great conversation. And in that conversation, one of the things that I asked, I said, hey, why don't we do more to figure out how we can bring the Indian kind of startup community and leaders like you who are really helping to shape, you know, the the next generation of companies, founders, ecosystems. And why don't we just trade thoughts? Okay. And you, you know, you actually brought a bunch of your colleagues and other investors came to Mountain View. Yes. And then you came back and then you sent an entourage of I think 60 founders came to Mountain View where we immersed in really the Bay Area just what was happening in the Bay Area from a technology perspective but also bringing in the knowledge the the experiences the business opportunities in India and just like having this kind of you know kind of different uh different perspectives mixing it up but you know anyway I just think it's like as we were chatting about, you know, this moment right now from an investor or building a company or being at a large company. You know, it feels very different to compared to pretty much anything. So, I'm curious to get your perspective on that. Yeah. You know, before we get started, Jay, thanks for being here and launching some of these extraordinary technologies to the community here. you know you know all of us as investors as the AI is starting you know it's like a trend I'm seeing is half the time we are kind of coding ourselves now so I've been building a a a markdown editor for in my cloud code so I can submit my articles directly into your product LinkedIn but I saw this part when can we get access to this well you have access to GitHub you have access to VS Code, GitHub CLI. The GitHub app, which I think is going to be really, really wonderful to use, is coming soon. So, the team is working literally seven days a week, day and night on polishing and finishing a bunch of really cool ideas. And it's fun to actually see what's happening behind the scenes with the team. It's a small team, right? This is not a big team. We have like I don't know maybe 15 people now 20 people on this team that you know on the CLI you can even follow the change log they're releasing a new version of CLI that you can update almost every day right with new features. So some of the things that we showed today there's remoting capabilities you can assign a fleet of agents you can do the research stuff. So yeah, I'll set you up after this. No, so this is u as most of the investors are becoming sort of coders and most of the coders are becoming angel investor there is a reversal of kind of things happening now because it's almost impossible. I think you'll be better at my job than I would be at your job. So uh excellent. So you know the the the way to launch in parallel access via CLI your research it's like so much power like I have like 100 hands to be able to get a lot of things done you know very very exciting you know last time when I had a session with Kyle when he was visiting then with you to now it's like a exponential difference in the suite that we are seeing it I think the whole community will get benefited sooner you release to everybody body. Yeah, one of the fun stories actually I forgot to say the GitHub copilot CLI. So what Karan and others showed today actually one of our fastest adopters for the CLI inside of Microsoft was not the engineering team. Wow. It was our legal team. Right. So actually that connected with some of the knowledge kind of word document, PowerPoint, Excel etc. they actually use this as their daily driver to actually get more work done faster and to reduce some of that toil that repetitive work that they would normally do. So to set the context I think you know at least I am in the an extraordinary situation of ability to create new products new innovations on top of these platforms that you're all creating. Um when you look at the last I would say 20 25 years internet to mobile to cloud to data waves after waves of technologies coming now the AI is like a gigantic wave in the center of all of them is at the end of the day the builders and the developers and as you said India is the one of the largest the large uh the largest developer er community and we are bringing more into the ecosystem and secondly the largest users of this I feel we are sitting on a an extraordinary opportunity for everybody in this room for building not only for India but for globally the innovations that we have never seen before and it's just the gap between what I call as being developers to being imagineers if they can imagine something they want to build with these kind of capabilities. I think we will see this innovation cycle much faster and that's what we discussed last time we had and now we are seeing that starting to happen in India. So we are in the center of gravity of what I would call a second wave of innovation. First wave of innovation of LLMs orchestration layer is all happened in the valley. We are seeing it. We are absorbing it. But the second wave which we'll talk about it is happening in India. Yeah. Absolutely. And one of the things that you know we've seen that journey internally even in with the folks that I know and I work with which is you know this in some ways like the folks that are staying curious about the technology right and you're staying curious you're trying new things out. Actually when I was doing another round table with a bunch of founders one of the questions that actually shocked me was this individual was asking like hey in a conventional company we would normally allocate maybe 15 maybe 20% for learning time in a normal company right so your average builder your average engineer would spend maybe 10 20 15% learning new technologies right in this age of AI he was noticing that his engineering team needed 50% 50 half of their time they needed to spend in learning and trying new things out. Right? So that was like a significant shift of learning time that was allocated here. But really it's, you know, it's in some ways it's not a tedious thing anymore, right? It's like if you just have the curiosity to go do this and you push yourself, you push the AI systems, right? And we talk about internally kind of this like camp one, camp two. And camp one users are folks that are using AI and they're most of the time they're like, "Whoa, this is amazing." They're mostly just astonished with what AI produces, right? You're like, "Wow, that's really cool." Right? And then there's camp 2, which is like you're perpetually frustrated because you're throwing really hard stuff at the AI and it's disappointing you. It's mostly getting it right, but you know it can be better, right? And what we all have to keep in mind is that if you're mostly in camp one and you're mostly just like, "Wow, this is amazing." Most of the time, you need to push yourself into camp 2. You need to be curious. You need to throw at it bigger, harder problems. When it doesn't work, try different models. Try to use different skills. Try to use MCP. Understand how to fine-tune the models, right? maybe RL like keep pushing yourself to hill climb on the technology because one of the things we all know is humans are really really bad at understanding exponentials. And as you said we are in the the exponential of all exponentials this century. So yeah. So I almost say that you have a low floor high ceiling product starting to show up. So entry barrier has been significantly reduced. I can start wherever I am and the product adapts around that. But as I build and understand and push the envelope, I can build very high ceiling products and go deeper and deeper and sky is the limit right now. So that capability is democratized to everyone across the globe and it's just how people imagine a world and how do I build together for that. It's like in the history I don't think this was ever possible and this is the first time anyone can start because the floor is been removed so low anybody can walk in. Yeah. And I think it's just also really important as you are building features or products whether it be you're running your own startup or you're working at a larger organization or you're in school you know it's really important to think through what that time to value is right because if you're building something and you want people to use it think about how they're going to be delighted with an experience within minutes right how many people here are patient right nobody Right? You want you want some gratification. You want some surprise. You want some delight in every product you use, right? And the ones that don't, you're like, "Oh, I'm not going to I don't I don't like it." You're probably not going to come back to it either, right? So, when we talk about things like taste and these things, it's not just like, "Hey, it has to look nice." It has to be this smooth experience that gets you to this moment of like, "Oh, there's things valuable to me, right?" And it's got to do that not after hours of setup and clunky configuration files and errors and weird documentation. It's got to be simple and then you can unveil and show more advanced capabilities, right? Because this is what's happening in the age of AI is actually I think even if you're trying to bring AI applications into the enterprise, so much of this has to be easy to to use. Yeah. So on that aspect, you briefly touched during your keynote on how SDLC of the old era is changing in the new. So if…

Transcript truncated. Watch the full video for the complete content.