Video: What Shipped at GitHub Universe: Put it to Work | Duration: 2740s | Summary: What Shipped at GitHub Universe: Put it to Work

Transcript for "What Shipped at GitHub Universe: Put it to Work": Hi, everyone. My name is, Kevin, solution engineer. Nice to meet you all. And here, we also have Alec. I'll let him introduce himself. Hey, everyone. My name is Alec Buchanan. I'm a solution engineer at GitHub working with customers primarily in the corporate space. And as Kevin mentioned, I'm joined by him, and we're really excited to talk about all the exciting shifts that we had last week at GitHub Universe. Yep. I know we still have some people. For me, it looks like we have a handful of people first off. Thank you all so much for joining. I just wanna give a few more minutes and then we can get started. Kevin, while we're waiting for folks to trickle in, should we dive into the poll? Sure. Yeah. Yeah. So we have a poll here. We like to ask it at the beginning of these sessions, mostly just to get an understanding of where people are. And it's really, you know, just for us to see how how things are trending amongst the audience, but also to help tailor our messaging a little bit to make sure that we're, you know, speaking to to you all, like, with stuff that's as relevant as possible. Cool. See some results trickling in. Looks like it's an even spread. Yeah. And regarding this question, it's really... This question has actually evolved over the past two years. I remember a year ago, we used to just ask if you're using AI at all. Could be Copilot. Could be anything else. But we've noticed that there are different things and different levels of depth and and and, the breadth of services that people are using AI across the, SDLC or AI DLC has continued to evolve so rapidly over the past twelve months. So, that's why these questions are now a little bit more, in-depth as far as, like, if you're regularly using it or, you know, to what extent you're using it. And we're gonna get into that as well during the presentation about, like, maybe some of the future facing things, as well as things that just shipped recently at GitHub Universe. Cool. That's really helpful to know. It looks like most people are are actively using it. So that's that's good. That's really good feedback and good to know. Cool. So we're gonna get started. First off, really appreciate you all taking the time out of your day. I know it's kind of like that that time right before the holidays and, you know, code freezes are are coming up in a number of weeks. For some of you, also, maybe you are taking, time off. So we're gonna get right into it. First off, you know, what is GitHub Universe? GitHub Universe is our annual developer conference where we announce all the new exciting features, and we have people from all over the industry get to one place in San Francisco to talk about all the great things around the developer workplace, workflow. And more specifically, it's been about how AI has impacted and and helped benefit the developer workflow and productivity. And so that happened last week at San Francisco. And during that, we announced 61 different chips. So we are not going to be able to get into all 61 today. We're gonna get into maybe four or five different features and workflows that we think are gonna be most relevant to the most amount of people. But I wanna call that out just because we do have a list of resources that we're gonna share around that recap and some of the things that we're gonna be referencing as well. So the big picture here with GitHub universe that for today is really just about addressing a lot of the friction point that we've seen across the SCLC. So what we're gonna explore is how development is changing, how GitHub is evolving, and we're also going to show what this means for your team, in general, but also specifically as it pertains to the GitHub platform. So I'm gonna quickly go through some slides here. Not gonna go through every single one of them, but I do have to go through them sequentially. But the big thing here that we noticed across many customers that we've spoken to is that development has continued to accelerate. Right? So teams are moving faster and taking more complex work. But as that occurs, we've noticed that there's been, more whole requests being created, more code to review. These forecasts have grown larger. There's a lot of unstructured, activity and and agents that people are trying to coordinate. So there's a lot of, like, institutional work that needs to be done, both from the customer side but also on the GitHub side, and we helped address that with some of the things that we're gonna show today. So I'm just gonna quickly go into four or five big things that we announced at GitHub Universe. The first thing here is that we are simplifying access to AI across your development organization. So today... As of today, GitHub Enterprise now comes with a access to GitHub Copilot without a separate subscription. If you are an existing enterprise, this is something that you will need to talk to your sales team to potentially opt into. You can still stay on that license where you have either GitHub Copilot Enterprise or GitHub Copilot Business. But moving forward, an option is there for everyone, who is a user on GitHub Enterprise to have access to GitHub Copilot, and you would just pay for the consumption of tokens across the different services. So this is really exciting. 90% of the features that we're gonna show today are gonna be ready to use out of the box without needing that separate subscription. Another thing that we're also gonna get into is what we're calling GitHub Copilot projects. So kind of zooming back, we announced something called GitHub... The GitHub Copilot app. This used to be just a wrapper around the GitHub Copilot CLI, but it's evolved to be so much more. It's where people are able to coordinate across multiple sessions, across multiple repos. But one gap that we've noticed is that when you're working in a project that has multiple repositories where code lives, where it needs to coordinate amongst those, it's still challenging to coordinate with that and make sure that when you're making a change in repo a, if it's related to repo b, how do you get that context in the right place at the right time? With GitHub Copilot projects, you can now coordinate tasks across both of these repositories where there are multiple repos, and you just focus on coordinating it from that project level view. We're also gonna get a little bit into Copilot automations, really just kind of a show of what it looks like. This is where you can do something akin to a cron job where you can schedule tasks to be, created or performed on a daily or weekly basis. And so when there are things or workflows that you're manually doing across your day, these are things that are right for automation, in that Copilot app or also in the UI. Another thing that we're also gonna get into is Copilot Fusion. So this just went GA recently, in Versus code in the CLI as well as the Copilot app. This is where you... Rather than picking an individual model in your GitHub Copilot service, you would just choose Copilot Fusion as well as the reasoning level, and Copilot Fusion will allow you to essentially have frontier level intelligence when it's needed for the task at hand. But then it will shift into more efficient or cost effective as well as faster models, when it's not needed. And so we have seen a huge amount of benefit on this, as far as making sure that you have that frontier level capability to tackle some of the most difficult challenges you have, while at the same time, reasonably getting below, like, 30 or even 40% below the cost of any of those frontier models. So this is really exciting, and we're also gonna show what that look... Looks like as well. Another thing that I will also get into briefly is AI scans for repository. So this is in public preview. Unlike all the other stuff, this is something where you can opt into it without needing to get through, you know, like an exception list. But this is where on top of CodeQL, which is our first party scanner for GitHub Advanced Security, which is an add on to GitHub enterprise to scan for first party code, dependency vulnerabilities, malware, as well as secret detection, you now have the ability to turn on AI scans to supplement those first party scans with, either languages that we don't cover natively or to essentially, bring in a reasoning model to identify some vulnerabilities within your first code. So I'm gonna switch over to my, local view. Just give me a second here. Let me get it from far. But let me go straight into the Copilot app and then show what that looks like. So I'd like to show here, is really just showing you what that Copilot project looks like as well as briefly explaining what Copilot Fusion is. But before we get to, you know, like, into the weeds here, what is the Copilot app? Right? So the Copilot app, this is where... It's essentially a desktop app for your team members, for developers, for builders, could be for product managers, could be for project managers, designers. There are a lot of different ways that you can customize this surface to work exactly the way that you need to. But you can see here that on the left side, there are a bunch of options to be able to view customizations such as MCP servers or maybe skills for agents that you can add in. You also have these individual chat sessions that have existed for the longest time. These sessions are... You... Are typically scoped to a repository just like a single repo, but we recently added this projects tab. And and what this project tab is is that when you are going to create a project, for example, which I'll do here, you can now either choose no repo if you wanted to work on something that is not necessarily scoped to repo, or you can choose one to many repositories that it can gather and coordinate across, within this context window. So think of it like these individual sessions here which were one off, you can now have this parent session, which is the project level view. And then in that project, you're gonna be able to coordinate task across, multiple repos. So I already have it set up here under virtual universe 26. This is a repository or actually two different repositories. One is a front end application, and then one is the API that is connected to it. And so the big challenge here is, for this example, it's really just about making sure that when you run something, you're able to coordinate across those two different repos. Right? But more importantly, something that I haven't shown yet, which is here in the chat session, this is where you're gonna be able to coordinate and see the different model view. But if you go to the top right here on this button, this is where you're gonna be able to get that extended view into individual parts of the project that you're working on. Right? So if I go back into overview, when it comes to looking at the work that's already been done or the sessions that are currently working, maybe open PRs or discussions that you're trying to work with, This is where it can all exist. And then when it comes to artifacts such as images, attachments, files, things like that, you can simply add it to the library here as a reference guide for the project to work with. If you have any automation, that's something that you can also see here. I don't actually have any, so it's blank. And what I have here, if I go back to this view, is what we call a canvas. So a canvas is where you can have, like, a diagram canvas or a browser canvas. It's really just there to be a visual reference for what you're working on. And what's really neat is that you can also click on pick and polish, for example. Click on something, it'll add that as a component in that chat window. So if you're working on something, for example, in the front end and you need to reference something very quickly, you can do things like that as well. So what I'm gonna do here is I'm going to copy and paste a prompt just for the sake of time. And what this prompt is, is it's going to look at this application here, which is a front end, which is, like, summarizing some of the different things that we shipped at GitHub universe in this, like, fancy UI. But the challenge here is that if you go to the keynote highlights, you're gonna see that there are a lot of keynotes. Right? Like, this continues to scroll. It can take a second a little bit. There's a lot of stuff here. Right? So I wanna be able to add a filter to it. That's what I'm gonna do with this prompt. I'm going to hop and paste this right here. And just to kinda go through it, I'm gonna ask it to build an interactive, you know, highlight experience across these two repos following these two issues that were previously created. I'm also gonna start exactly two coding sessions scope to each repo. And then, you know, just kind of adding some additional things here. Now I'm gonna enter this just for the sake of time, but what I'm also gonna show as that's running is you might have noticed that instead of picking a particular model, if I uncheck this, you can see all the different models that you have turned on such as Gemini, Opus, SONNET, OpenAI, all those different models are available. But if you turn on Copilot Fusion, you instead are gonna be able to see these three different effort levels. And so this is what I wanted to kinda cover as it's running, and you can notice that the first task is running with OPUS 5.5. What Copilot Fusion does is that it is essentially deferring the selection of the model to Copilot Fusion. Right? So it is in a model orchestrator that lives on top of the Copilot harness, and what it does is that it will do one of three things. If it is a simple task, it will just take a simple single model. That is akin to what you saw with auto mode where it will choose and route the task to a particular model. So if it's something that is straightforward, it'll do that. But if you are using Copilot Fusion, whether it's, higher max or standard, and it's something sufficiently complicated, it will potentially go into a cascade mode where it will do a very initial first pass. If it doesn't feel like it matches the task, it will then upgrade itself to a more powerful model, and it'll continue to do that until it matches the task. Another, modality that it can do as well is what we call critique. And with the critique version, that's where it's gonna use two different frontier models. So for example, OPUS and Sol or OPUS and Astra, and it will use that, one of them to create the task and then the other one to critique it so that it can, essentially rubber duck itself and then make an overall higher quality, output. So the thing that matters here is that we have seen... And and for what it's worth, I think we added this in the in the, resource guide. In a third party analysis from artificialanalysis.ai, there's that Pareto frontier where on the x axis is the cost of an individual task, and then on the y axis is the frontier level reasoning. We've seen that Copilot Fusion Max is within two points. So I think it was, like, low seventies versus low seventies. Two points up Opus 5.5 max as far as Frontier intelligence capability, but at half the cost. And more importantly, it's also even half the the time of the times. You'll find that when you're using it, it's really, really quick. I highly recommend trying this out. And another thing I'll also mention is that it does adhere to your model policy. So if you have an enterprise owner who is setting the industry models that they have access to, that's something that it will, honor as well. Last thing I'll call out is that you might have noticed here, there's also a 30% discount on this until November 30. So try this out. Really great. Could not recommend anymore. So you see here, you know, going back to it, it did that... I mean, pass with Opus 5.5 to, you know, free up those ports, do the initial thing. We're also seeing it do that main pass now with GPT Luna, so it's now dynamically switching between some of these models. And kinda like what I mentioned earlier. Right? These are individual task, and this prompt has multiple things that it's working on. But it started with that single model, now it's using cascade to essentially find that right model to to pair with. Show sole. And you're gonna see here that it's gonna eventually start implementing features. But if I go back to this overview tab on the top right here, you can also see now from an orchestration view, you can see these two sessions were created. Right? So I can go here and see all the things that it's working on, what has been completed in the past. You can also see it in that overview here. And the big thing here is that you might notice, looks like it's walked over. It's already implementing some of the features. But if we go back to the overview tab here, there's gonna be a sub session, a trial session here that is in autopilot mode, with Copilot Fusion standards when it's selected. And so if you are wanting to work and coordinate across these sessions that are now being kicked off, you'd use this session on the bottom. But this bottom right here is to steer that active session that was created by the project. And so now when you're working across multiple rebuilds, you can do it all within one pane of a class. It's really, really powerful, and we're gonna see it continue to build out, some of the different things. One thing else I wanna show is maybe the other part too. As you can see, like, one out of three, it already implemented that front end preview, which is why I was, like, flopping back and forth just to show us what it did. And you can see here, same thing. Right? It's doing that same cascade reading files, essentially accessing it. But we're also... If I go back here, you can see it's starting to build up that keynote highlight. Yep. One thing while this is running that I also want to show just because, we just mentioned earlier, our automations. So anything here, just keep in mind, it doesn't need to be scope to repo as well. Right? But it can be... It... You could also optionally choose to just make it something where... Maybe it's a generic task. Right? Could be something like emails or it could be, you know, doing outreach for customers, that kind of thing. Think about, like, anything in your daily workflow that you could automate. It's very much, very likely as possible here. And so I actually have some of these automations here just to show. And you can see here that there's a list of these automations, whether they're idle or currently running, the previous runs. And the cons of these are not super, super important. But what matters here is that if I choose to edit one of these, you can see that all of these are easily customizable. You can choose the schedule day and I have this Chrome mic one where it works at 5PM my time. Have these different prompts that it's gonna run. You can even have the automation use a particular agent or a particular model, as well as a mode. So if you want to essentially build something that you need to run on a daily or weekly basis, you can create an automation directly within the GitHub Copilot app as well. Really cool stuff. Not gonna wait for this to completely show, but I just wanna refresh. And I just refreshed it here. You can see that now there's this keynote highlight section. So it's already almost done building it. It's really, really quick given the fact that that prompt was, like, five paragraphs. And we can see here now that it's already done, some of that. Now what I wanna do is, shift over to the github.com experience to cover some of the things around Copilot code review and co... Specifically the max level reasoning that we've also added as well as the AI scans. We do that right now. Oh, accidentally went into that. So, just give me a second here, and I'll just cover this in a bit. But this is that same repository or, specifically, this is the front end repository that was scoped to that previous session that I was kicking off, in that in that project view. And this is an existing pull request. The first thing that I wanna highlight is that this is the new view. You may have already seen this already. You can choose to switch back to the previous one as well. It's like an option at the bottom. But one thing that we've noticed is that there are a lot of comments. Right? A lot of things happening with automations from GitHub actions, GitHub Advanced Security, GitHub Copilot, individual comments, just kind of like constant scrolling. Right? So how you address that is we consolidated the things that you need to see as far as, like, passing checks, approvals, and resolved comments here. And then if you need to actually go into the comments, there's now the ability to filter and show the specific things here. But the the main thing I wanna highlight is that there is now in GitHub Copilot code review the option to add, the max level reasoning. And so if we go to the reviewers here, this pull request has already been reviewed by Copilot review at max. One thing I do wanna highlight, there is this individual benchmark for what we call review dash bench dot a I. And so this is a third party independent, benchmark where it produces an f one score, which is the matrix between the precision as in how, many, you know, results it actively finds, and then the recall, which is how many of them it actually finds out of the one better, total. And we have found that, Copilot ClearView Max is, like, near the top. I think last I checked, it was third. So we are now seeing that it has one of the highest f one scores. And out of those two, the precision is, like, super, super high. And so what this means is that when you are choosing Copilot code review, whether it's light balanced or max, you are now getting pretty close to the frontier level as far as, like, capability. We also recently updated light to now use an ensemble of agents to review each other to old... Overall improve the quality of the work as well. So if you're still using light mode to get those, like, sub ten minute reviews, we've also increased the precision and reduced the cost of that by 20%. But I just wanna call that out is that now the salience of the findings that you're gonna get as far as the peer review as well as the coupon code review, have gone up meaningfully. And what that looks like is now in this new view, if you go to the Copilot comment, from Copilot code review, it's really gonna give you that summary first off. Like, here's all the open issues. Right? Here's the severity. Here's the link to it. But more specifically, now that you have this, you know, pull request has been created with this review after max, or this comment that's been reviewed, how do you actually go through to address some of these things? Now what we have done is we've also added the option to add things to batches as well as ask for clarification to Copilot. So one of the things that we've gotten feedback on is like, hey. All these are helpful, but I just wanna make sure that I have the ability to either batch it all together into a single large commit if it's something that I wanna do at scale, especially as we get today, we get more and more comments, but also be able to ask you questions. Right? And so if we go through some of these things that is it recommending, you can see that there's the option to add it to batch and also ask it the question. So if I click on ask a question, it's gonna open that chat window that refers to that, and you can actually ask it using Copilot and picking a particular model to get feedback to it. Or you can just ask it for clarification, maybe modifications as well. You can also just choose to add it to a batch where it's gonna continue to append these comments and then add it to a single commit. So lots of options here. And the reason this matters, by the way, we've seen that when people are using the Versus code IDE extensions for Copilot or just in general adopting Copilot across their, their local environment, the size and the volume of pull requests has just exploded. There are so many pull requests that people are reviewing now, and so what naturally happens is that there's a bottleneck, at the pull request level. And so wanna call that out. One thing that I also wanna mention is that if you want to automate the creation of pull requests, review in the settings tab at the repo org level, you can also create a branch rule set that targets a specific branch or subset of branches and then make it so that it automatically, calls Copilot code review as well. So, what we have seen is that with Copilot code review, the customers that turn it on versus off during evaluations, they've been able to get the number of days that appears open to merge, from about nineties to five. So a 40% reduction, and that's gonna help increase that throughput and make it so that you can deliver software faster and at a higher standard. The last thing I also wanna show with the time that I have here is some of that stuff around the AI scan. So so far, we've been talking about that local environment with Copilot projects in Fusion and with Copilot overview max now being an option directly in the CCR experience. But if I switch over to this findings tab here, you might have noticed that this is a PHP repository, which we don't, we don't, support out of the box with CodeQL. Right? So this is right for something where we could run that AI scan to start finding some of those vulnerabilities. This could also include shell scripts, could include languages, that we don't cover as well, like Groovy or Apex, things like that. But what you're gonna notice is that you can go into tool here and see that under CodeQL, you can also see that there's AI scan as well as DeepAI scan. So AI scan is a public preview. Deep by AI scan, which I'll talk about briefly, is in private preview. The AI scan, what that does is it will run Copilot to essentially nondeterministically scan for some of these CWEs. And so similar to what you see on CodeQL, you're gonna get that here where you can see the CWE, the line, and then that recommended fix. But more importantly, just like with CodeQL today, you can assign these fixes to Copilot. Right? So if you identify these issues just like with CodeQL, you can assign it and do that bulk. One thing I'll also briefly touch on is when I go back here, deep AI scan. So this isn't private for you, but I just wanna talk about this really quick because it's here. This is using Microsoft's, which is the parent company of GitHub, Microsoft's code name, Amdash, to scan essentially at a deeper level. So it reads for your code essentially like a very skilled security researcher. Think of it like using a frontier level model, in the back end to find a ton of different vulnerabilities. Compared to some of the Frontier AI labs, we've been able to see that the findings are are sometimes, especially when you do, like, multiple passes, able to outmatch in terms of the f one score at a fraction of the cost. So definitely wanna call it out. This is gonna be really promising for doing that deep scan. But, you know, just one thing I wanna quickly show with the time I have here is that if you do assign that fix to, an agent, it's gonna create a pull request similar to what you saw earlier. You can see this is generated by Copilot. And rather than needing to manually address these vulnerabilities, you can now use that agent auto fix. It's gonna spin up an ephemeral runner with the minimum MCP service that you specify. And, essentially, it's gonna go through it, reason through your code, fix it, verify that it fixes the issue without causing problems. They can do that several times to do multiple passes. And, really, this goes from having to go to your development team, maybe to security team, to go through that process of remediating individual vulnerability, which we have seen takes about two hours. We've seen that this can cut it down to about fifteen to thirty minutes. Right? And that's for bulk fixing where you're just now delegating that challenge part and then just simply reviewing the work that's done in the pull request. So I know that was a lot that we just kind of mentioned. You know? Once again, really appreciate your time. We were only able to cover about four or five different things today, But, hopefully, that was a good overview into just the high level strokes of, like, what we're doing and how we're envisioning the platform because it's kinda scaling over time. But with that being said, we do have fifteen minutes for q and a. Alec and I will will help with anything. And if you'd like to chat with us, we really appreciate your feedback as well. And, hopefully, that was helpful. Great presentation, Kevin. Really like that. What are some of the features that you've been most excited about since post Universe launch? I know there's too many to mention. I think I went through the change log yesterday. I had 31 different tabs open. So, curious maybe if there's anything that we didn't cover today that, maybe is, like, a sleeper feature that you're really interested in. Yeah. Let me think for a second here. I think we're making a lot of investments into not just the individual features, but also the scale of the platform. And I I can't articulate it as well as they did at the GitHub universe recap, but the number of pull requests and the number of volume of traffic that we've seen has gone up seven x year over year. And, as you... You know, I've I've all seen that that has a... It had some impact on the, the stability of the platform as well, and we're making huge investments in partnership with Microsoft and Azure to help scale that out. I don't have anything super concrete. All I can do is just mention that, you know, at the recap, we did spend a a meaningful amount of time there. And that's really just to make sure that in this era where we're seeing developers increasingly lean on multiple agents across multiple surfaces, the stability of the platform as well as the, capability that is gonna be able to more than meet that need as well. Yeah. Alec, do do you know any of the specifics around that and some of the volume numbers? I don't. I know some of the traffic that we've been seeing just in the last few months alone has has surpassed essentially all the commit traffic that we've seen, to this point in GitHub's history. So, really, if you tune into the keynote, you probably saw the the graphs that we had on stage, and, really, the the commit traffic is just exponential onto the platform. So, I'll show that same sentiment as well. I'm really excited about the investments that we're making into the infrastructure side of the house. Yeah. Yeah. And, something I did not really touch on today mostly just because of the, maybe not, like, the the time constraint, but more so the the demo ability of it is a workflows. So we just g a that. That's been a public preview for the greater part of the year. And what that is is now you can use, the Copilot SDK directly in your, CI steps. So within GitHub actions, and that is right for just the most amount of creativity. We've seen some really, really cool things from customers. Could be anything like a self healing CI. Could be things like automatically creating, assigning issues or pull requests retroactively, doing health checks, that kind of thing. So highly recommend if you're, you know, living in the get of actions world to take a look at the agentic workflows if you want to embed reasoning. Not to, you know, replace the manual steps and rigor that's required to build something and make it ready for deployment, but more so to invest some of those recent things. But I know, Alec, you're you're much more, familiar with Git of Actions and I. Would love to hear your thoughts on that as well. Yeah. I think the GenTech workflows is amazing. You know, historically, CI and actions, was very deterministic. Right? Like, you had to explicitly state what you wanted to do, when you wanted to do it, and then tell it how you wanted it to run. But with the Genentech workflows now, you can explain a workflow that you want in natural language. You can then go ahead and add that workflow to your repository, and then it's able to do some of these nondeterministic things like triage incoming issues, analyze bugs, and route those to the appropriate teams. There's a whole new world out there that, people can be using, not only just the Genesys workflows, but GitHub actions as well beyond just your typical CI workflows. Yeah. Absolutely. I see we have a question in the chat from Jose. So the question is, for Copilot enterprise orgs, can Advance restrict which models Copilot Fusion is allowed to route to and set spending limits per org or per team? How do they interact with existing model policies? Great question. So today with Copilot Fusion, it does honor your model policy. So if you are restricting particular models that are in scope for the usage of Copilot Fusion. So for example, in my presentation, they use Opus 5.5. I believe it also uses Luna as well as Sol. If you have any of those turned off, it will not ever route to those. And and also to kind of dig in further, if you happen to be on GitHub enterprise with data residency, which is our cloud deployment that is, you know, scope to a specific, stamp within Azure, that will also honor things like the data residency model that you selected as well as the FedDev compliant models. And then as far as the budgeting is concerned, you can set spending limits today, by individual user. You can also do things by a particular cost center, and then you can scope it out to a group of users as well. And so there's a lot of different ways that you can slice and dice that spending, control. Could be enterprise wide, could be in... By user, could be by cost center, which is scoped to a group of users. So hopefully that, answered your question. Let me know if there's anything that I missed, Alec. Yeah. I just wanted to call out. Simon had a similar question around being able to determine and, sort of configure which models Fusion was able to use. So, actually, in the enterprise policy settings now, you do have a separate section for fusion specific model selection. So you can go in there. And if you have a model that's generally available for folks to pick manually, you can actually go into fusion and turn that off. That won't be present, infusion selection process. And I think there is a world in the future, I'm not making any promises, where potentially you could have models that are turned off for everyone in the enterprise, but then have fusion able to access those. So, just thinking ahead kind of where we're going with this. But, today, now you do have those controls in the enterprise. Yeah. Great question. That's, something that we know, especially when it comes to governance and and managing these different models. I I think it's kind of like a double edged, sword, right, where we are this open provider where you have with one subscription access to five or six different providers. And, there's just, like, so many different policies that change, like, literally week to week. I think we had one week a month ago where we released three different models, like, announced at the same time. Like, Luna came out and Sol came out and then Opus 5.5 came out. And I remember, a lot of customers were trying to manage, like, all those at the same time. Have a good day. See a question from Simon. At what exact point can it change the model? Is that my impact cash is right? That's a good question. And it actually depends on the reason level as well. So with Copilot Fusion, there's that standard. There's the the the high as well as the max. High and max are more similar to each other than standard is. Standard is, under the hood is actually the auto mode that we've had for a while where it will just pick the single model, and then high and max is the one that can choose that particular, path. So what I think would be helpful is if we share the doc on Copilot Fusion to help explain, like, how it it does it exactly. I don't know if we have, you know, an exact time or, like, the order of operations as to how it it it routes it. The Allegino, that is something that we we provide, visibility into. I don't know off the top of my head. I know that Fusion does try to choose a pattern for each prompt to route along the cache boundaries. It also tries to keep the main conversation on the same model model where possible to benefit from that caching, but I don't know the exact underlying mechanisms. I will say, I was using a Copilot project session earlier this week and a little bit later last week where I was making some changes across about five different repositories. I was using Fusion High for about two to three hours, and I think my session total was about 500 AI credits. So I thought that was pretty token efficient considering the output that I got. Yeah. Yep. Yeah. That's, my experience as well. I I had something running for two or three hours, and it used $6, which is observed. Right? Like, you know, it's actually, like, building stuff. It wasn't just using input, tokens. Yeah. And I'm actually gonna read the doc right here because, hopefully, this this will help answer it. So Copilot Fusion also considers real time system health and availability when choosing models. Routing occurs along natural cache boundaries and large changes in conversation complexity to avoid additional cache related costs without quality improvements. So that... I know that that doesn't kind of go through every single edge case as far as, like, model a to b or model a to c. But but hopefully that that helps with, at least, you know, providing some guidance as to how it, adheres to those different cache boundaries. I will add if you're seeing unexpected behavior, if you have a GitHub account team, reach out to them, provide that feedback. We are in very close contact with the, engineering product and design for Fusion, so we can give that feedback directly to them and be your advocate. We also just wanna make the product experience, a great one. Since Fusion's come out about a couple months ago, I've already seen it improve leaps and bounds, and, I have no doubts that it'll continue to improve as the months and weeks go on. Question from Simon. Any way to limit credit limit on the orchestration session? I know that we have slash goal, and that's an alias for autopilot. And and with goal, you can specify the number of turns or the number of tokens that it can consume. But I don't know if that's, like, one to one to exactly what your, your question refers to. So if I I don't... Do know if we have that, but, do you know if there's anything related to that, like, as far as, like, in individual session, Al? I'm not sure at the top of my head as far as that level of granularity. I don't wanna misspeak, but from the top of my head, I I don't... I'm not familiar with that. Yeah. Yeah. Typically, our our, budget controls are more so scoped to, like, end of calendar month. Yeah. But that's a good question in order to, like, maybe bring that in. Generally speaking, if if you know, like, as far as those, like, cost versus Frontier capability is concerned, oftentimes, we'll see that the standard is is able to perform, very well for 90% of tasks. And then, I I would say high is kind of our standard go to at least when we do do our work, internally. We sometimes use max when we do, like, plan mode as well. Yep. I know that we also drop some docs, in the chat. So we do have, like, six different resources. A lot of that is referring to thing that we mentioned, like, the GitHub universe recap. The Copilot app projects is really helpful for understanding how it works. GitHub Copilot now, you know, being included in GitHub enterprise. That max review effort, AgenTic Autodix. So those are offered there if you want to dig deeper into some of these particular discussion topics. And then I also, we also added the fusion benchmarks as well. So that's gonna be that artificial analysis, benchmark where you're gonna see that frontier, Pareto... Is that called like the Frito? I don't remember what it's called. But it's the one where there's, like, the curve where it shows, like, how expensive it is for a task versus the capability. And we're in that quadrant where it's, under a certain dollar amount, but then above a certain threshold for quality, which, we're really, really proud of. Cool. Any other last questions? We're... If not, we can, you know, include here. Really appreciate the engagement from from the audience and, you know, with, just kinda being on top. It's it's super helpful. But, yeah, thank you so much for for joining, and and we hope that that, was helpful information to understand, like, where we are today, where we're going moving forward. Some of the things that are GA today and then regarding security, some of the things that are are gonna be, you know, in the horizon, both publicly and private preview. Yeah. Thanks so much for tuning in. Thanks, Kevin, for an amazing presentation, and we will see you all on the next one. Take care.