Fixing merge conflicts and PRs with Copilot cloud agent | GitHub Checkout

I'm seeing a thread here. Keeping developer choice with Copilot coding agent, you have so many options. Welcome back to GitHub Checkout, Tim. Thank you. It's fun to be here again. I think every time on GitHub Checkout, we kind of say like, "How long ago was it that I was last here?" And and can't remember. It's probably something crazy like 6 months now and and so many things have changed in that time. What I want to share today is a a bunch of improvements we've been making to Copilot Cloud Agent, which used to be called Copilot Coding Agent until last week, which is an agent that, as the name suggests, runs in the cloud. So, you give it a task, which you can do from a bunch of places, including github.com, VS Code, the GitHub mobile app, and lots of other places, too. And then the agent works in the background in the cloud, and then you can come back and review its work once it's done. So, here I am on one of my repos, and we've got the agents tab, which was added, I think now maybe 3 months or so ago, but it feels like forever. And the agents tab is really your kind of HQ for agents within GitHub. So, you can come here, and you can send new tasks to agents to work on, and you can keep track of things that your agents have been doing for you. And what you see at the top here is just a simple prompt box where you can ask the agent to do what you want it to do. By default, that starts with Copilot, but we also have the Claude and Codex agents available, too, and the model picker, as well. So, when I'm on the Copilot agent, I can pick from a bunch of different models from OpenAI and Anthropic. So, I can pick the model that I want to use that I like best or that I think is the is the best fit for my task. So, all I have to do in here is type in what I want the agent to do. So, as an example, I'm going to say, "Implement rate limiting, allowing only 60 requests per IP per hour." I'm going to go with Claude Opus cuz I like that, and then I hit the start task button. Then Copilot's going to get to work on this in the background. If you've used Copilot [clears throat] Cloud Agent before, until very recently, as soon as you clicked that button, Copilot would open a PR for you, but we've actually changed that a bit recently. The reason why we we aren't opening a PR immediately is really two things. We wanted to allow you to work with an agent in private, rather than for being forced to like immediately shout to all your team like, "Hey everyone, I'm working on something." Like, instead, you can now work with an agent, figure things out, and then decide to open a PR when you're ready. Just like you would do if you're working in VS Code or with the Copilot CLI locally, or if you were just writing code yourself. And also, similarly, or or kind of separately, allowing you to use the agent for things that aren't just writing code. Something we heard from lots of people is they wanted the ability to ask the agent to plan before implementing, and that was kind of weird if you got a PR straight away because you don't really want a PR for a plan. And similarly, being able to do deep research. So, ask an agent questions about the code base, let it work in the background for maybe 5 minutes, 10 minutes, 20 minutes, and then come back to you with an answer. And again, if you're just asking a question, you don't need a PR for that. You just want logs, and that's what you can get. Got a bunch of kind of examples of that. So, the one we just started asking it to implement rate limiting, it's working on that. It's doing this in its own environment in the cloud where it can kind of run bash, run tests, all those kind of things. Here's an example from earlier where I asked the asked the agent a question, and here it just gives me an answer to the question. It doesn't force me to open a PR, force me to write a bunch of code. If I want, I can reply and say, "Let's remove that." Or let's add that. Or let's Or like I can move forward to a natural conversation. In this one, I asked Copilot to implement or to create a plan for implementing rate limiting. It comes up with a plan and tells me what it's going to do, and then I can ask it to go ahead and start work on that from the plan. Really nice workflows here that allow me to have the agent working in the background in the cloud, but not just be limited to coding, but allow it be able to do planning and research and other kinds of things. And to have control myself over when the pull request gets opened, as well. So, I can decide when I'm ready. So, going back to that example we were just looking at, this is another task that I gave the agent earlier to implement a change. I can see it's finished, and at the bottom, I can easily jump into the diff, so I can see what the agent's done. And when I'm ready, I can hit the create pull request button and actually get a pull request for my team to review. But until then, it's private. It's just for me, and I can decide when I want to move forward with that change and ask my ask my colleagues or indeed ask Copilot for a review. But I can take it from research to planning to execution to PR in the same session? Yeah, that's right. It's all just very fluid in that respect. If you think about your agent in VS Code or if you're using the Copilot CLI, it's not really like tied down to one kind of thing. You just put in a prompt, and it does what you say. And we've really tried to make Cloud Agent feel exactly the same way. You could start with a question, "Does this repo implement rate limiting?" You could then pivot that into a plan and say, "Well, create a plan for implementing rate limiting." And then when you're ready, you could say, "Okay, now go ahead and do that implementation." And that can all be in the same thread and with the same context, as well. So, the agent can remember the stuff that you already said, rather than you having to kind of keep pasting in the same context again and again. The same way as you can use the agent in VS Code or use the agent in the CLI, you can also use an agent in the cloud to do pretty much whatever you want to do. That's really good. [laughter] Okay. And then of course, moving on from that, here I've got a PR, and I can still interact with Copilot in some of the same ways as I've been able to do before, except we made things even better than they were in the past. This is my PR that I just created with Copilot, and you'll note here that it says, "Mention Copilot in a comment to make changes to this pull request," which is something that I really, really love. I can jump into the code that Copilot has changed, and I can ask it to make whatever changes I want with a comment, just like I would do with a human reviewer. Let's say I spot some test coverage that's missing. I can just ask Copilot to make a change, or if I if I just don't like something. So, here for example, it says it set an expectation that the body length should be greater than 100. I can say like, "Copilot, let's make this a specific number instead." And then Copilot can jump back in, pick up the pull request, and make changes. So, that's how you could do before, but the thing that I'm really excited about is that we've made this experience much, much better in human PRs. I have a pull request that I created myself. This also allows me to mention Copilot in a comment to ask it to make changes. Now, before when I did this, the workflow wasn't the best because Copilot would open a pull request on top of my pull request, and I'd have to merge that in. Some people liked it, but some people didn't, and we've now made that much more smooth. Now, I can say, this example, the tests are failing, so I'm going to say, "Copilot, fix the failing tests." And instead of the agent deciding to open a PR on top of my PR, it's just going to come in, contribute straight to my branch, and make those changes, which is way, way better, way, way more fun. But I will add that if I want it to open a PR, I can still do that. I just have to ask for it. So, I can say like, "Copilot, create a PR to fix the failing tests." And instead, it's going to follow that old PR stacking flow. One thing that you kind of see here is that we're just trying to make the agent smarter and more flexible. I love it. You can have it your way. And one other one I want to tell you about that I don't have a kind of demo pre-prepared for, but I think people are going to love is that this flow of mentioning Copilot also works for fixing merge conflicts. I've been getting a lot more merge conflicts on my PRs recently. Like, with lots of agents running at the same time, I'm merging more code, I'm making more changes, and I just get a lot more conflicts. And maybe it's bold to say this, I don't think any developers like see a merge conflict in the merge box and say, "I love that. That's my favorite thing as a developer." Most people hate doing that. It's really painful. It's really fiddly. Well, now Copilot can actually do that for you. Copilot, fix the merge conflicts, and Copilot's going to jump in and fix them for you. The reason that works so well is because Copilot in the cloud has its own development environment, it isn't just making the changes and hoping that they work, it can run the build, run the tests, run the linter, and make sure that after fixing the merge conflicts, the stuff actually works. When Copilot comes back, you're actually in a good state where things are working well. So, I don't have to make the changes, edit them, and then hope and pray that it works because Copilot's already doing all of that for me. The way that we always think about Copilot, like we want it to do what a great developer on your team would do. I learned something new today. There's too many things. I think like last [laughter] week we had like six entries in the GitHub change log with just features, features, features. So, it's always hard to keep up, which is one of the reasons I love coming on here to get to talk about things and and show things. One more thing I want to show actually, as well. Um when Copilot pushes, so like when it opens a PR, or when it pushes to your PR, by default, you have to click a button before tests run. And that's kind of a security thing. It's to make sure that someone has reviewed the code before tests run, which can be important if your actions has secrets in it, for example. You'd want to review that and avoid the chance of like something weird happening to your secrets and being leaked. We've added an option that allows you to customize that if you want to. So, if we go to the settings for the repo, and then go to Copilot, and then go to Cloud Agent, we now have this tick box where we can say whether we want to require approval for workflow runs. We have warnings here, and the default is that you will have to click that approve button before the CI/CD runs. But you have the ability to turn that off if you want to and say, "Actually, I'm happy to take that risk, and I want the CI/CD to run immediately." Sometimes that could be nice just cuz it speeds up the feedback loop. You get that green tick or the red X more quickly, so you can move forward faster, rather than waiting for Copilot to push, then coming back, then clicking the approve button, and then waiting for the feedback. You can get a faster feedback loop this way. Oh, I love this. I'm turning this on myself cuz there are things that that don't really know. Like, I know what the workflow's going to do. There's nothing sensitive. If I'm updating my own personal website, like it's fine. Maybe in like a enterprise context, you want to have those approvals cuz there's like maybe secrets, API keys, stuff in there. But in my public open source projects, for example, I don't have anything like that. Like, I just want Copilot to be able to work as fast as possible. I want the best feedback loop possible. And then the next step that we're going to be working on, which I'm really excited about, is if the CI/CD fails, then Copilot can jump in automatically and fix those. Instead of waiting for you to like come back and say like, "@Copilot, fix this." Letting Copilot come in and and do that for you. I think that's going to be super super nice and I'm I'm very excited to to get to that point, you know, getting self-healing PRs. Copilot does have its own development environment and in general it is going to run your tests, so you hope in most cases that when it pushes it's not going to get a red PR, but there obviously are going to be cases where it can't run all the tests, the same as I can't on my own computer. But what's magic is if Copilot can come in and fix those automatically when they do fail, that's going to be an even better experience cuz what I want to get to is just green PRs, quickly approve them and quickly merge them and then move forward. So that's what's next. Anything else that's coming out that you can tease out? We have a bunch of validation tools that are built into Copilot and we now have the ability to turn them off if you want to. We've had a lot of improvements in the mobile app. I think another area that I'm really excited about right now is that we're working on a public API for Cloud Agent as well. That will create new ways that you can start these tasks, new ways that you new ways that you can track them and I always love APIs cuz they just let developers be creative and create things that we hadn't even imagined. With Copilot Cloud Agent you have so many options. And also what you mentioned earlier about the sessions being always on, now an API, I see where we're going and I like it. Working wherever you want, working however you want, with whatever agent you want, whatever model you want, that's what we're all about. Amazing. Well, thank you Tim for showing us everything. This is very exciting. Can't wait for you to come back in a couple of months. I hope so. Thank you very much, really appreciate it. And that was your brand new look at Copilot Cloud Agent, now more flexible and smart than ever. Check out Agent HQ and use the agent that's going to help you do the task at hand, be it Copilot Cloud Agent, Claude, or Codex. Thanks for watching. If this was helpful, please don't forget to like and subscribe so you never miss another developer update or tip. Push those changes to main and I'll catch you on the next release.