The Agent Awakens: Collaborative Development with Copilot - Christopher Harrison, GitHub

Allright, well, let's—uh, let's get started here. I did see a couple people walk in at kind of the last minute. If you haven't already read the slide there, you'll notice the slide has the first bit of instructions. Send an email.

That email address that you seeright up there: pec.john+aiworld@github.com. And then, just as the subject, just put in your GitHub handle. One quick note about the GitHub handle: if you are using Enterprise-managed users, this is going to need to be a personal GitHub handle.

So that email address, and then your GitHub handle as the subject. Again, not—but you know what to do. So go ahead and do that. And while everybody's getting settled in, let's go ahead and get rocking and rolling. So, my name is Christopher Harrison.

The gentleman over there is John Peck. And we are both Enterprise Advocates at GitHub, obviously doing a whole lot of developer relations stuff, so basically talking to customers and doing events like this, primarily focused in on developer experience in this little thing called GitHub Copilot.

So this is going to, hopefully, be a hands-on lab. Now, I am going to put out a couple of real quick caveats because we had the exact same lab this morning and we ran into a couple of challenges, the biggest of which with the Wi-Fi network.

Now, because what we're going to be doing is going to rely quite heavily on the network, we're going to try, okay? I'm going to repeat that one more time: we are going to try. We are going to give it our best effort.

If it works, wonderful. And if it doesn't work, here's what we're going to do instead: you will still have a day or two to be able to complete the lab, and you can go ahead and complete that on your own time.

And what I will do instead is I will start doing some demos, I will answer a whole bunch of questions. The morning group had a whole bunch of questions, and something tells me that this group will also have a whole bunch of questions.

Been doing enough Copilot talks, there always seems to be an awful lot of questions. More than happy to talk through all of that. So, we are going to try our absolute best. Hopefully everything will work. And if it doesn't work, we've got a nice little fallback.

And I want to make sure, again, that I highlight: you will have access to the lab for a couple of days after this event, so that way you could go out and do it on your own time from there.

So, one last time, because I did see a couple of people that just walked in, notice instructions on the slide. Please send an email: pec.john+aiworld@github.com. The subject, just put in your GitHub handle again. Make sure that's your personal or a non-Enterprise-managed user account.

If you don't know what Enterprise-managed users are, don't worry about it, because then you probably don't have one. If you have one, you'll know it. I see a couple of knowing nods there. Okay. Without any further delay, let's go ahead and talk a little bit about GitHub Copilot.

So John's going to be working in the background here. Once he is done working, I'm then going to go ahead and make an introduction and get everybody pointed at the lab. But I want to get started by talking about something that's very, very important when we're working with GitHub Copilot, and that is context.

You'll notice that we have a tagline for GitHub Copilot called "Your AI Pair Programmer." Now, not only is it wonderful marketing—it kind of is—but it's also a fantastic description of how best to think about GitHub Copilot. That you want to think about it as an AI Pair Programmer.

And the example that I always give, the analogy that I always give, is my partner and I woke up the other day and she says to me, "Hey, I want to go to brunch." Fantastic. Who doesn't love brunch?

And so I recommend that a spot, and she looks at me and says, "You know, here's the thing. Been there an awful lot lately. I want to go somewhere new." "Okay, fantastic." And then I recommend that another spot.

And then she says, "You know, now that I'm thinking about it, I'm really kind of craving waffles." Spectacular. Who doesn't love waffles? And then I recommend that another spot. Now, that's a very normal back and forth. Now, what happened was, is a request was made, I had some level of context, and I gave back a perfectly valid response.

And then some more context was provided. And then I gave yet another response. And that's a very common back and forth. And that may have been context that she had at the very beginning and maybe just hadn't yet shared.

And it might also have been context that she sort of discovered as we were talking. Again, all of this is very, very normal, and it's how we think about when we're talking about with other people, when we're talking with human pair programmers, and then also in turn with that AI Pair Programmer.

That context is key. Now, when we're talking about context, a lot of times we really focus in on the prompt when we're talking about context and AI. That I see an awful lot, especially with new developers, to Copilot, is they're really worried about, like, "Okay, well, how do I craft a prompt?"

And don't get me wrong, prompt crafting is an important skill. But context goes beyond prompt crafting. Because unlike a lot of other tools, which are maybe going to run all of your code through different parsers, or maybe even build your code, or potentially even, like, run your code, Copilot, again, behaves like an AI Pair Programmer.

It's going to be reading your code. Just as if I brought somebody else new onto the team, and they're going to start reading your code. So making sure that that code is readable becomes very important. So if you have good function names, good variable names, if your code is well-structured, Copilot's going to be able to navigate that pretty well.

And if you're using a lot of single-letter variables, if you're using a lot of abbreviations, Copilot's going to come along and it's going to struggle with that, just again, as another pair programmer would. In addition, comments. Now, we all know the importance of commenting code.

And there is a strong part of me that has this feeling that good code doesn't need comments. But here's the thing, is having a line or two of comments does help you make your code that much more readable.

And so even if it might be superfluous, if it's going to make life easier, maybe just go ahead and add that in. Because it's not going to, say, slow down your application. So having comments, describing what it is that you're going to be doing, how it is that you're going to be doing it, is really going to help things out.

And then finally, last but not least, is your project structure. That having a good, navigable project structure is also going to help Copilot. Now, the real takeaway that I want you to get out of all of this is this is sort of like just good code 101.

That having good code is really going to help out an awful lot with context. I'm going to come back to questions a little bit later. Okay. So how do we then help Copilot help us? And so number one, context, we talked about that.

Number two, the intent. Make sure that you're clear about the intent. What is it that you are trying to do? And be clear about that intent. Now, one of the things that I see an awful lot of developers do, especially if they've maybe worked with, like, older chatbots, is when they start sending those initial prompts, that they're very terse.

Because they're used to dealing with older chatbots that maybe weren't really good at processing natural language. And so it basically became like using command lines. That's not what this is built for. So be clear. If you want something to be done a particular way, and kind of getting into the specificity aspect here, tell Copilot.

Or as I like to say, don't be passive-aggressive with Copilot. That sometimes I'll see a new developer, they'll go in, they'll give, like, a very short prompt, and then they'll go, "Well, wait a minute, why didn't Copilot figure this part out?"

Well, did you know that part? Well, yeah. Did you tell Copilot? Well, no. Okay, we'll tell Copilot. Like, if there's an important piece of information that you need Copilot to consider, go ahead and make sure that you tell Copilot that.

Okay. Now, Copilot supports a handful of different workloads. You'll notice the one on the far end, Copilot Coding Agent, I'll get to that one in a minute. It's all bright and sparkly and so forth. That's not because that one's necessarily best.

It's just because it's the newest, and as we all know, we need to just highlight and sparkle the newest item. So, first up, let's talk about code completions. So code completion is sort of the original version, if you will, of GitHub Copilot.

This is where I'm going in, I'm writing code, and then Copilot will suggest the next line, it will suggest the next function, it will suggest the next class. And this is wonderful. If you're in the zone, you know what you're doing, you're typing away, fantastic.

And Copilot will be able to support youright there. Next up is chat. And originally this was single-shot. That I go in, I ask a question, and I get a response back. So maybe it can generate a single file for me.

But the problem with that is we think about even the simplest of updates. Like, if I go to update a web page, I'm going to be updating an HTML, JavaScript, and CSS file. That's multiple files. And so this is where edits come into play.

And with edits, Copilot is now able to edit across multiple files. And then we take that to the next level with agent mode. And with agent mode, now again, still inside my IDE, what Copilot's going to be able to do is navigate your project, figure out what needs to be done, and then start building code.

And it's able to work iteratively. And it's able to run external tasks, like, for example, running your unit tests. And if it detects, "Hey, there's an error," it's even able to self-heal. I'm going to sneeze.

Thank you. Okay. I was waiting for the second one. I usually sneeze in twos. It'll show up in a minute here. Okay. And then finally we get into coding agent. And coding agent is different because this is going to be done inside of github.com.

So I'll have an issue on github.com, and I can then assign that to coding agent. And coding agent will then work in the background. Figure out what needs to be done, be able to again run scripts, be able to self-heal, and then eventually create a pull request that you can then review and eventually merge into your code base.

And so this is what we're going to be exploring. I'm not going to worry too much about these two slides here. There we go. This is what we're going to be exploring in this lab. John, how are we doing?

We are doing splendid. I have two people left to add to the board, and then.

I will need about three minutes to add them all to Copilot.

Okay. But we can at least go in and start accepting invitations.

You can go in and start accepting invitations. One quick note, Michael Pope, I don't believe I have your correct handle. So if you could send me a new one, that would be great.

Okay. And there was one person who walked in just a minute ago, so we'll need to get that sorted as well. But we can get there in a minute. Okay. So, first part to the lab. Navigate to the URL that you seeright there.

So github.com/octoholhandsonlab. So O-C-T-O-H-O-L, Hotel Oscar Lima.

And then at the very top, and by the way, I'm assuming that you're logged into GitHub with the handle that you sent over. And then up towards the very top, you should see "Accept Invitation," and then go ahead and select "Accept Invitation."

Now, here's the most important part about today.

Silence means, "Hey, everything is fantastic." So if you're not seeing it, please let me know.

Not complaining today.

Ah.

Sorry. We'll see.

Okay.

And then "Accept Invitation"right at the very top. Okay. Yep.

Did you send the email?

No, I didn't.

Okay. Allright. I'll come back in a second.

Check this box, ask Christy.

Hey, John?

Yeah?

Do they need to ask for the Copilot seat?

They don't need to explicitly ask. I am just going through the processright now of adding those in.

Okay. Perfect.

If you forgot to take the button to ask, go back and do that again.

It doesn't matter.

Okay.

Yeah, he's going to be adding you in. Okay. Now, for anybody who did not send the email address, please go ahead and send an email to that address. And then with your GitHub handle, I'm going to leave that up for about another 15 seconds here, because I'm trying to keep things moving for everybody as quick as I best can here.

I can share the slide deck. My handle is GeekTrainer, and you can basically find me on everywhere with that, including GitHub. GeekTrainer. G-E-E-K, Trainer.

GitHub?

GitHub. At GitHub.

Okay.

Now then, let me get everybody rolling here. So, after you have accepted the invitation, who has not yet accepted the invitation?

I don't know. The email didn't go through or something. I feel like I sent it.

Okay. And then what happened when you went to OctoHol? And by the way, I want to make sure that I highlight something. You'll notice that more people raised hands. Silence means everything is fantastic. So if something is not working, please let me know.

Now, for people who everything is working fantastic, then here's what you can do, is go to that URL. Actually going to modify the slide real quick, is

and if we could get phones to be silent or vibrate, that would be wonderful. Okay. Allright.

So,

to get started with the lab, what you're going to do, navigate to that URL that you seeright there. So gh.io/agent-sdlc-lab. When you get to the part to create your repository, the owner is going to be octohl, and then the name is going to be your handle, again, replacing your see where that's going.

And go ahead and start working on exercises 0 and exercises 1. And I will let you have at it, and then I'm going to start going around and helping people get unstuck.

Yeah, same issues now with the email. We haven't had an accept invite.

Okay. Hey, John?

Yeah?

When you have a minute?

Absolutely.

Okay.

Right over here. Okay.

Did you send the email?

Yes.

Okay. We'll go ahead and get John over to help you out in a minute. Yes?

I sent the email like this, and I haven't.

Okay.

And

I can't see anything.

Did was it maybe there earlier? Do me a favor. Hit plusright at the top here. Hit new repository. Just let's just test something real quick.

Hit the dropdownright here. Okay. We'll go ahead and get John over here in just a minute. Sorry. Yeah. Okay.

Allright.

Now then, remember, silence means everything is fantastic. So if something is not working, please raise your hand. Otherwise, we are never going to know.

Oh, okay. So note the owner. So owner is going to be octohol? Nope. So notice the slide. Okay, sorry. Yep. And then the repository name is going to be your handle. There you go. Yep. Hey.

I didn't see the repository name yet, so.

Okay. Then hold on for just a minute. John will be over. Hi.

Oh, okay. So here's what we need to do. Hi.

Yep.

Nope. So, yep. Perfect. So go ahead and hit enter.

And so go ahead and scroll down. And you see the step 0 here, set up the environment. Let's go ahead and just hit that. And then now we're going to follow the instructions from here. So you'll notice that you're going to want to open up that URL in a new tab, and then you're going to select use a template and then create from there.

You're welcome.

Allright.

You just go directly to github.com/octohol.

Allright. How's everybody doing? I know there's a couple of people that John's going around to. Yeah.

Can we create a copy of the template that the owner needs to be octohol?

Yes. Yep. Beautiful.

Who else is having a trouble getting added? Allright. I sent the email off. I just sent it.

Oh, okay. Doesn't matter.

Living on the edge.

Yeah.

Nope. Here's what you're going to do, is so go to oh, okay. So go to docs,

and then just scroll on down and go to step 0. And then now go ahead and followright along with what this says here.

Yeah.

Oh, okay. So nope. So we need to go to the first part here. So navigate to that URL here.

So go to.

Allright. Who else is needed adding?

Yep. So real quick, as people are getting started, I want everybody to kind of notice the steps on the slide here. So step one is go here. What you're going to find there is a set of instructions for the hands-on lab.

And then in step 0, it's going to have you then create your repository. When you get to that step, you're going to set the owner to octohl and then the name to your handle. So start here, start working through the lab, and then you're going to go in and set that up.

Okay. So let's go into step 0.

Yep. And then now you'll notice it's going to have you open that up in a new tab. It's going to have you create from.

Create from GitHub.

Exactly. Yep. And then you're all set.

Don't sit down yet.

Yeah, so start at 0.

Oh, okay. So here's what we need to do, is let's go ahead and let's just go to

nope. You see the URL up at the very top there, the gh.io?

Yeah.

Okay. Nope. It looked like we didn't it should be sdlc. So let's go ahead and just try that again. So type in type, nope, nope. Stop. Yep. There we go. So gh.io

slash agent

dash sdlc. Nope, nope. Yeah, you have an s-l. We want sdlc. Yep. Dash lab. Enter.

Software development like Cyclops.

There we go. And so now go ahead and scroll on down. And then you'll notice there's step 0. Let's go ahead and goright into step 0. And then now it's going to have you create your repository. So if we read through, we'll notice step one.

We'll open that URL up in a brand new tab and then follow the instructions from there.

Okay. So.

Here?

Yep.

Nope. Create a new repository.

Yep. Nope. Put it intoright as it says up on the slide. Hit the dropdown. Octohol. And then the repository name will be your handle.

You can leave it as private. That's fine.

Okay. Allright. You're welcome.

And then you'll see the results.

Yeah.

I created a brand new repository. Not opening it. Can I delete this?

Yeah. You can just go ahead and delete that one. And then, oh, okay. So all you have to do is go to settings.

Scroll all the way down. All the way. Yep. There we go. And then hit that. And then you're going to hit delete. And then you're going to hit delete. And you're going to you can actually just copy and paste it.

And then, yeah. Welcome.

Christopher.

Yo.

There's no explicit instruction for them to merge the PR after it was created.

Correct.

So leave it?

Yeah. We can leave that unmerged.

Allright.

Yep. Yeah. Yeah. We're just going in and kind of exploring it. There's also

yeah, I'll get to that in a minute. Ask me that question again in five minutes. I don't want to interrupt everybody, but there's an important thing there. So ask me that question again in five minutes.

Yeah.

Cool.

I visited this. I opened that specific lab, but I don't see use template entered. Maybe because I'm not signed in, or?

Oh, yeah. You need to be signed in. There you go.

Yeah.

Perfect.

Yep.

Nothing else,right?

Correct.

Yep.

Yeah. And just kind of keep on following on with the lab from there. What this is going to have you do in a minute here, is it's going to have you open up your repository again, and then you're going to create a couple of issues and start playing around with coding agent.

Yep.

There might be some code spaces where the actual Git connection hasn't established yet, because that happens later after the initial load.

Yep. Yeah. Well, the lab actually says just send a hello to kind of wake it up.

Oh, okay.

Yeah.

Great.

We had play for that.

You can't hit all of that.

I know. Keep rolling.

Yeah.

I've done step one.

Fantastic.

So I ran into something that I can't do. I run out of premium requests.

Oh. Hey, John.

Sure. I'll bring a laptop. There's two classes,right? There's a phone problem.

He's having a premium request error.

Ooh. I'm trying to find something. Okay. So I

made a request against your instruction results. Fixing a typo.

Oh, thank you.

I type really, really fast. I don't type accurately, but I type really, really fast. See, here's the thing about being a developer. You don't have to spell correctly. You just have to spell consistently.

Yeah.

Yeah.

And you can rely on that by the syllabus.

Yeah.

It's fine.

Yes.

Yeah. Actually, if you're in a zone and you already finished out step one and want to start getting into step two, you can go ahead and do that. You'll also notice that all of the steps actually include little optional exercises.

So if you are somebody who kind of just moves really, really fast, there's other stuff to do in that lab. So yeah. So go ahead and kind of keep working through exercise one. I'm going to do an overview of exercise one in just a couple of minutes here.

Yeah. Sorry.

For code spaces, no. So code spaces is a separate product. The reason that we're using code spaces is it's effectively a container in the cloud, and so it just eliminates setup. So it's wonderful for just development in general, but especially something like this, having people install software, forget it.

So yeah. Christopher, that's okay.

I also had the same question. How much is it that you get to each repository? Does it come out of the box?

So Copilot for Enterprise, which is going to be like everything that we're using, is $39.99 a month. So $40 a month. Yeah. There's also a business tier, which does not there's a few different features that it doesn't include, but one of the biggest is going to be access to a coding agent.

And that's $20 a month. Or again, $19.99 or whatever.

And that's before the Enterprise?

Yeah. Yep. Yeah. Yep. Yeah. Yeah.

Yeah. I'm just talking about setting up your code space. Is there a different.

Hit refresh real quick.

That's on?

Yeah.

Do you still have oh, there it goes. There we go.

Allright. So here's what I'm going to do. I know that people are still working. Periodically, I'm just going to come up here and I'm going to kind of talk and highlight a little bit about a feature. So this is basically a real quick overview of what it is that you're doing in exercise one.

You can have a seat. Make yourself comfortable. No, go ahead. Pretend like you're an adult.

Do I look like an adult?

No.

Proof that you work with somebody for way too long.

No. So let's talk a little bit about GitHub Copilot Coding Agent. So the whole goal of Coding Agent is to have

an AI peer programmer that what I can do is I can assign it a Copilot issue or, sorry, a GitHub issue and let Copilot begin work on it. Now, the first big thing about assigning an issue to Copilot remember that whole speech we did about context, waffles, all that good stuff?

Yeah, yeah, yeah. So context. Context is key. And so if we take a look at the issue that we created, we went ahead and gave it a pretty decent amount of detail about what it was that we wanted to have accomplished.

So you'll notice that we're going to be creating a functionality to allow for submission of games. So here's all of our requirements and kind of list them all out. And if this was going to be something more robust, I might have even dropped a potential workflow that I wanted it to kind of step through as it was building out all of the project.

Again, don't be passive-aggressive with Copilot. That if there's some piece of information you want it to see and understand, absolutely 100% tell it about it. Now, the next thing is, after you assign it, it needs an environment in which to work.

And

there is a workflow that you can create, and it's called, he said, "Speaking Slowly to Stall," Copilot Setup Steps. And so when Copilot goes to do its work, it's going to be running inside of GitHub Actions, and it's going to step zero run everything that it sees inside this workflow.

And if you look through this workflow, what you're going to notice is a whole bunch of steps to install different services and packages and so forth. And that's now going to be the environment in which it's able to do its thing.

Now, that also leads me to another very big point, which is about security. Because here's what I'm going to do. I'm going to take an AI agent. I'm going to assign it a task, and it's now able to go off and start doing things.

And I want to make sure that everything it's going to be doing is going to be completely contained. So the first big thing about working with Coding Agent is that it does not have access to the internet. That by default, it is a firewall that's going to block it from any external resource.

Now, you do have an ability to open up the firewall if you do want it to be able to access external resources, but by default, it won't. And that includes, by the way, installing frameworks and libraries. So the only way it's going to have access to frameworks and libraries is either A, you open up poles in the firewall, or B, you set this up to install all the necessary libraries.

Number two, this is running inside of GitHub Actions. And if you're not already familiar with GitHub Actions, one great thing about it is that it's going to launch an ephemeral space. So it's going to set up a space.

It's going to do its work inside of there, and then it's going to collapse down from there. So it will not have access to any other resources besides what's inside of there. Number three, it does not have write access to your repository save for the branch that it creates.

So it cannot modify anything inside your code base. The only thing it's able to modify is just its branch. Number four, that branch, you're going to notice, will not actually create a full pull request. It's going to leave that in draft form, not become an actual pull request unless you come along and say, "Yes, I approve this."

Number five, along those lines, if you are a good developer shop and you have different requirements, things like security checks, things like unit tests, end-to-end tests, regression tests, all of which need to pass, what you're also going to notice is that on this pull request, none of that can actually run

where's that button? Right there. Until you click that button that says, "Approve Workflow Runs." So again, trying to keep it completely self-contained. And then finally, last but not least, and this was a question that was asked earlier about reviewing it and then merging it in.

When it comes to the current version of Coding Agent, there is a restriction, and this is by design. Remember, again, Coding Agent is brand new, so this can always change. But this is by design. That if you assign an issue to Coding Agent, you are not allowed to then review its work.

So you need somebody else to review its work. So basically, it's sort of like in most organizations, you're not allowed to review your own work. You need to bring somebody else in. Sort of the same thing is that I assign the peer programmer, and then I need somebody else to come along and say, "Yes, that this actually is OK."

Allright. So questions on Copilot Coding Agent.

So let's say you are the only programmer,right? So would that accept this creator?

That is a fantastic question.

And I am just going to say that that is a fantastic question.

And I'm going to say that's a fantastic question.

Other questions.

Can you iterate on what it generates?

That's a great question. So the question is that was also a great question. So the question is, can you iterate? And the answer is yes. What I can do let me zoom out a little bit. That was a little obnoxious.

Is I can just go in and add a comment. So if I go in and I add a comment and I say, "Hey,

what I potentially want it to do, also add in an endpoint to search for games by name." There we go.

And note here he's adding this comment to the pull request, not to your name.

Thank you. And if I scroll back up, give it just a second here, and then we'll notice the cute little eyeballs again. So now what's happening is that Copilot has started to work again. And if I hit View Session, hopefully I've stalled just long enough here, you'll notice that it's actually started up a brand new session.

So it's going to take the work that it did, and then it's going to start back up again and begin work again from there. Yeah.

Does it work with self-hosted runners?

No. Not at the moment.

When?

I shouldn't have even said at the moment.

Let's summarize that answer. No. Yeah.

I always feel like a press secretary whenever I'm doing this.

Can you run Copilot on behalf of other organizations? If you have a GitHub app and you have other companies that integrate with it, can you run a GitHub Copilot on behalf of them, or do you have to just run that yourself?

So

give me a little bit more about what you're asking there.

Yeah. So I have a GitHub application. Clients install my GitHub application. It generates enriched prompts that we want to use with basically GitHub Copilot to run something on their code base. Can we do that?

No. So what you could do is you could stand up an MCP server so that way they could use Copilot, talk to your code, and then start to interact that way. But you're not going to be able to have Copilot do something to somebody else's code base.

OK. So it's not like an open API that you can just have a Copilot stick in a code base and do stuff?

No. Yeah.

It's like the same question here. We have a GitHub app, and we're creating a pull request in the customer's account. So I think we cannot leverage.

So yeah. So what you can do is if you have a pull request, you can assign that to Copilot at that point. But again, it needs to be on a repository that you own.

OK. So if we install the app, so it's of course in their repository, and we can technically assign to Copilot to do some work on behalf of the client. Yeah. Right?

I don't know the answer to that. I'm going to be honest. I don't know the answer to because I don't know necessarily what an app is able to do

on behalf of the user at that point.

We should try this.

Yeah. I would need to get back to you on that. Yeah. My handle is GeekTrainer. Let me just

there we go.

So just shoot me an email, and I can try and dig up a more robust answer for you. Other questions? Yeah.

What if I'm the salesperson in the coding department? I'm already doing the pull, and then how do I add it into production? Do I add it as a comment? Yeah. What if the company basically not what I want?

Yeah.

Yeah. This is a great question. I'm actually going to uplevel this a little bit. So when we're talking specifically about Coding Agent, if it doesn't do something that I wanted it to, or if it does get something completely wrong and keep in mind that AI, just like human developers, is going to periodically get things wrong, and it's sometimes going to get things spectacularly wrong.

Again, just like humans are going to sometimes get things spectacularly wrong. So if that's the case, there's a few different approaches that I could take. So one of them would be is just like I did here a minute ago or three minutes 57 seconds ago, but who's counting is just add in a comment and just clarify, "Hey, this is really what I want to use."

So that's one option. If it's completely wrong, if it's just completely out of left field, then the next option is you can just go back all the way up to the very top here and under Assignees. And I'm not going to remove this because I want to let it keep doing its thing.

But just unassign Copilot, change your issue, and then reassign Copilot. And then that way it will just completely start from scratch. Number three, and I think this is always very important when we're working with AI, is if we see it get something wrong, one of the first questions that we really want to be asking is, why did it get it wrong?

And nine times out of ten, and I'm going to comeright back to this word yet again, is context. Is that it didn't fully understand not only just your request, but the environment in which it was going to be working.

So one of my absolute favorite features is

Copilot Instructions. So on Copilot Instructions, what you're going to notice, and this is read in Coding Agent, and it's also read on every single chat request. Oh, how do I

yes. I figured out a shortcut key on a Mac. Go me.

Thank you.

How's your copy?

Hey.

What you're going to notice here is I have copilot-instructions.md. This is looked at by Coding Agent, and it's also looked at by every single chat request that I make inside the IDE. And what you're going to notice in here is I have a really nice high-level overview of what I'm doing and how it is that I'm doing it.

So I've given it a little bit of introduction about the app itself and how it is that I'm building it. I've given it code standards. I've told it everything that's required before commit. I gave it some global language guidance, and then I gave it some specifics on what it is that I'm building and how it is that I want it to be built.

And then if I scroll all the way down, you'll even notice that I highlighted that repository structure. So that way I'm clearly spelling out, this is where to go find things. So this is a big place to go in and look if you want Copilot to have some additional context, some additional information.

Another big thing that you can do, again, speaking to things that both Coding Agent and Copilot Chat are able to use, is to stand up an MCP server. So this is actually one of the other exercises in the lab is Model Context Protocol.

And basically what an MCP server allows you to do, and I have a slideright here that shows you the diagram. I need everybody to memorize this diagram. There's going to be a quiz on this later. Fantastic. But basically, here's what happens.

If somebody goes in, they make a request to Copilot, Copilot can then look through the list of commands that have been exposed out through different MCP servers, and then from there go access external data. Because this is always one of the challenges whenever I'm working with Copilot or really anything that's built on top of an LLM, is that typically it's going to be limited to just that training set, and that's it.

This allows it to access external information and even perform operations on your behalf as well. So if you see Copilot Coding Agent doing things in a way that you weren't quite expecting, take a look at its context. Take a look at what it has.

So look at your Copilot Instructions. Maybe put an MCP server. So if you've got a really large repo or a really large code base, and you want Copilot to be able to better search and better reason over that, this is a place where standing up an MCP server would come into play.

Now, I do want to highlight one last thing because Harold is in the room, and he would be very disappointed in me if I didn't highlight this. And this is .instructions files. And I am not saying this just because Harold is in the room.

This is one of the things that Harold works on. I'm not saying that just because he is here. This is true. This is one of my absolute favorite little features where when we think about what we're doing in coding, we're doing a lot of the same thing over and over again.

I'm going in and I'm creating yet another React component. Or in this case here, I'm creating yet another Flask endpoint. But I probably want all of those to follow a particular pattern. And so what I can do is I can create a .instructions file, and I can then manually add this into chat.

And this .instructions file can then have information about how it is that I want it to be created, the different rules around it, and then even point to prototype files that basically say, "Hey, you see that games file over there?

Use that. That's how I want this to be built. Hey, you see that test file? Use that. That's how I want it to be built." And I can even have this happen automatically that you'll notice that I have apply to header info.

And so now if it matches this pattern, so if it's inside the server test folder and starts with test underscore, it's going to automatically apply that. So let's say again, going to React components. I could then go ahead and set up a filter so anything that ends with a JSX, TSX extension, I want it to apply this.

If you're a Svelte developer like I am

allright. OK. Allright. I got a couple of people. Fantastic. Beautiful. Then I can go ahead and set up a pattern so that anything that ends with .svelte is going to follow the instructions here. Now, again, this part hereright now, only available inside of Copilot Chat.

But again, other ways that I can start to provide context. That was an awful lot on context. There was a question over here. Yeah.

How often do you update this as part of your software development cycle?

Yeah. This is a great question. Is how often do you update this? Keep in mind that this is an artifact of your repository, and I very much want to treat that as such. So I sort of give this an analogy to you know those little scripts that you spend like three days writing, but it's going to save you five minutes a day?

That's basically what these are, is that you're going to spend some time, and you really should, create your instructions file, create your .instructions file. Really make these robust because what you're going to wind up with then on the back end is increased productivity, but you're also then going to help ensure that the code that's being generated is generated the way that you want it to be generated.

So as things change, then do that. Yes, Mr. Harold.

Yeah. One of the things I need to add is space. You want to correct it. So let's say you get it on the. And you start minimal, and then you keep adding. And you see it here as an option to say, don't do that.

Or if you need to do this, do this instead.

Yeah. And I would even add to that that one of the things that's relatively unique about being in DevRel is I do an awful lot of demos. And one of the things about doing a demo is you need that to work.

And doing that with a probabilistic tool like GitHub Copilot does become a little bit of an adventure. That will be the word that I will use there. And instructions files have been probably the number one thing that I have leaned on to help ensure that I get some level of consistency.

So I really can't overstate the importance of instructions file. And again, I'm not just saying that because Harold's in the room. Yeah.

Can John have an opinion here?

Yeah.

Which is also as important as who is upfront. Yes, you have the person who's actually operating on the problem, but team leads, team managers, et cetera. That's part of your responsibility set. And as much as your team makes a decision about what set of dependencies we're going to use, your team's making decisions about what our patterns can be, et cetera.

And you're making architectural decisions. Those should go into instructions file when you make those.

Yeah. Absolutely. Yeah.

So in your issue, you specified you want to test DNS, but it seems like that would be something that would be embedded within an instruction file to ensure that there are always testings. Is there any reason that it's fear of a question of both, or is it a matter?

So the question is that, hey, inside your issue, you told it something that it probably should have been able to find elsewhere, that it was inside the instruction file, and you still typed it out. Do you really need to do that?

Fantastic question. So I'm going to answer that directly, but then I'm also going to uplevel that as well. So to answer the question directly, I really didn't need to. But this really, again, kind of falls under that category of not being passive-aggressive with Copilot.

If there's something that I know, typing it out isn't going to hurt anything. So adding that to that, if anything, it's just going to help it because now I'm just further reinforcing the fact that, hey, this does need to actually work.

Now, I am going to uplevel real quick because I think this also then brings up another question, which is, hey, now we've got the issue here. And especially if I'm working with Copilot Chat, where now I could have a .instructions file.

And we've already seen that I can have those call other files. And then now I've got my Copilot Instructions. And so that then kind of begs the next question, which is, well, is there a hierarchy here? And the answer is no, that there isn't any form of a lastright, that it's going to consider all of that context and then make its best decision from there, which also then kind of leads to another thing, which we should really be focused in on, like John was highlighting with our team leads, is let's try to be consistent.

So let's not start sending mixed messages. Because just like if we start sending mixed messages to our team, and they're going to get confused, the exact same thing is going to happen with Copilot. Yeah.

Can just one other question. Sorry. On the MCP diagram you had, you mentioned that you can't go outside.

I can open up holes in a firewall.

So that's specifically as a part of the onboarding of the MCP server, you have to open up the firewall as a part of that integration?

So if I'm using MCP locally from inside VS Code, good to go. If I'm getting Coding Agent to do that, then when I add in the MCP server, I also need to open up the firewall. All of it's actually really well documented to highlight all of that.

But yes, the straight answer is yes.

So many of these MCP servers use different protocols, so adding that configuration becomes a little bit difficult. If it's just WARP-based versus username password-based, it's easy, but then otherwise it's confusing.

Yeah. So when you go in and configure that, you can do that at the repo level. I believe you can do that at the org level as well. And then what you do is you then set variables. And you can do that again at the repo level, or you can do that at the organization level.

So you can centralize all of that.

Can you review the code that it's fixed on, and does it start a whole new session versus limiting just review this file and make use of this file?

So I'm going to inject my own question into that first, and then I'm going to come back and answer that question. So the first question is, can you review the code? And I want to make sure that I mention this because it's so incredibly important.

AI does not change the DevOps flow. That if I think about how I wrote code before AI, I would write code off my own memory. I would be writing code based on code that I'm seeing elsewhere. I would write code by copying and pasting from Stack Overflow, making a couple of changes, and then just crossing my fingers and hoping and praying that it works.

Or maybe that's just me. Or maybe I'm just the only one willing to admit it.

So in that situation, what do I have? I have a DevOps flow where I'm going to get John to come along and review it. I'm going to have linters that are going to come in and check to make sure that I'm following all my standards.

I'm going to have security scans that are going to look for any potential vulnerabilities. And I have all of these different things, all of these safeguards to help ensure that the code that gets into my code base is secure and is written the exact way that we want it to be written.

When I'm using AI, all of that still stands. So I want my manual code reviews. I want my security checks. I want my linters. I want my unit tests. I want all of that to go in and run.

Now, to get specific to the second half of your question there, which is, hey, it's just this one file that I want you to look at. When I'm working with Coding Agent, or really Copilot in general, there's always a non-zero chance that it's not going to follow what you want it to do exactly.

So if I say, hey, go in and update this file, it might work specifically on that file. And really, in particular, when I'm working with agent mode, it might work with that one file, but it also might do a couple of other things as well.

And there's not necessarily a way to guarantee that it's only going to do just that. If you want it to only ever do just that one particular file, this is where edit mode comes into play. This is where ask mode comes into play inside my IDE.

Because in those situations, I'm steering rather than letting

Copilot steer.

So when you review the code, it's going to start a whole session, and then you might get a different code with your comments not fully taken into account.

So there's always a chance that when again, Copilot is probabilistic. So there's always a chance it's not going to give me the code exactly the way that I want it to. But remember that at the end of the day, what this has done is this has created

just open this up inside of a brand new tab. This has created a brand new branch. And so I could go in and add in a comment and tell Copilot, hey, I want you to fix this. But this is also a branch.

So I could then just go in and say, hey, let me create a brand new branch on that space. Or I could go in and then just clone that, and then just check out that branch, and then go in and manually make the changes as well.

And so in the situation that you're describing where, hey, maybe everything looks good, but it's just this one file that I need to modify, honestly, in that situation, I think what I would do is I would just go ahead and

I would just do this, is

get checkout.

Yes.

You mentioned that it's like you need to go through every last detail of every Git.

Command.

Copilot's great at generating Git commands.

It is.

So yeah. So I could just go in and do that. Now I can just modify it manually. Or I could just work with Copilot here then on whatever it was that I was wanting to do. Yeah.

Yeah. So can you go back to the diagram of the MCP?

Sure.

That has the big kind of your computer section to that. So if you're using one of these remote agents and it's connecting to MCP, can it do that, number one, with the remote session?

Yes.

How long does that session look like on your behalf?

So I'm going to kind of uplevel all of this. So the question is, can it go to a remote server? How long does that session last, et cetera? And the short answer to can it access the remote server?

Yes. Can it act on your behalf? Yes. How long is that going to last? That depends. That when you go to set up that connection to that remote server, you're going to provide a credential for that. So in the case of, say, the GitHub MCP server, then I can go in and I can provide a token for that.

But then I can set an expiration on that token.

Yeah, for sure.

Yeah. So all that's really going to depend. There isn't anything in Copilot that's going to automatically sign you out or anything like that.

And so this MCP server, it can't basically kick off this is something that the actual agent that's running remotely has to decide to pull from. It can't basically kick off the agent itself.

Correct. Yeah. So the MCP server is not just automatically going to be able to do something on my behalf. I'm going to have to indicate, hey, I want to do this. And then on top of that and I don't think I've got it set up on here, and it would just take me a couple of minutes to do it.

But if I go in and I make a request that's going to kick off an MCP server, I actually have a button that says, hey, do you want to do this? Which actually leads meright into my next point, which is that I want to make sure that I'm vetting my MCP servers just like I would anything else that's going to access external information, just like anything else that's going to be able to act on my behalf.

This is really important. Can I ask one follow-up?

Sure.

Is it possible to create a GitHub issue on behalf of another company and then have that be assigned to their internal code?

We're getting into a lot of specifics. Let's talk afterwards.

OK.

Allright. Yeah.

Do you have visibility as to what model the Coding Agent uses or visibility as to how it's making decisions so that you can further prompt it in the future to help it make better?

So for Coding Agent and again, just to make sure that we're clear, this is where I'm on GitHub.com. Right now, we're not exposing out the model that's being used internally. And there's a reason for that. Because we have a very specific type of workload, a very specific set of scenarios that need to be accomplished here.

And so the model that we're choosing is the model that specifically that we've determined is best for doing this. There might come a time where that starts to get exposed out, and maybe that allows for change. But as ofright now, it's specifically because it works with this.

Yes, you still have your hand raised.

One more question, quick question.

Yeah.

Really good question. Is it possible to have more than one Coding Agent?

Yeah. I mean.

Like a pair of them, like can you make them sign up in the future? I guess it would cost more.

Yeah. I could create 100 issues and assign all 100 of them to Coding Agent.

So you could just do them even smaller.

Yeah. So remember that each time that you assign an issue to a Coding Agent, it's going to create an environment specifically for that issue.

So each issue will get its own agent.

Exactly.

Oh.

Yep. Yeah.

So trust,right? Apply to section.

Yep.

It seems like it's going to grow so huge. Can I do it just as opposite? In that folder, I've got my own instructions and stuff.

Overrides.

Yeah. Kind of like a chain and hierarchy. Yeah. Instead of having one gigantic file.

So

what you can do and I actually have an example of thatright here.

Oh, geez.

Yeah. So you'll notice thatright here, I'm actually inside this .instructions file, I'm actually referencing out to another .instructions file.

OK.

Yep. And then on the apply to, you could go in and start setting folders as well. So however it is that you want that to be set up.

OK. That's good.

Yep. Yeah. Yeah.

Allright. As you know, the context for MCP is not only tools,right? It's possible to access resources.

Sure.

So is that possible to configure the GitHub Copilot to access some internal reliability resources? We need all these structures. For instance, I'm working on a big enterprise where I need to get a pattern for code. Patterns for everyone else?

Yeah. So if you're working with Coding Agent, you can configure your list of MCP servers at the organization level or on individual repos

for anything that's going to be inside of VS Code. And I'm not sure how others would necessarily have that set up. You can actually just set up inside of a repository or inside your repository your MCP JSON file, which is what I've gotright here.

And then I can just list them all here. That now becomes part of the repository. The user will then still need to set up their own credentials because, again, it's going to act on that user's behalf.

But this is not one particular,right? I can't access resources in.

Correct. Correct. Yep. Yeah. Mr. Harold.

So Coding Agent only supports toolsright now. Resources just landed into this code.

Oh, OK. Thank you.

And then what am I talking tomorrow?

What time is your talk?

Oh, it's on the MCP track. You won't miss it.

OK. Yeah.

We have to worry about space collisions at all? Like MCP tools having the same kind of advantages and the agent getting confused about it?

So you know I'm actually going to kind of uplevel this. One of the things about using MCP servers all up with agent mode is that Copilot does need to figure out which MCP server to call and when to call it.

And generally speaking, it's going to be able to do that on its own. But again, and let this maybe be the number one lesson that we pick up from today, that if I ever see Copilot not doing something the way that I expect it to do, what am I going to do?

Write some instructions.

There we go. Or even more generically. I used to be a tech trainer. So I would literally just go around different companies to teach classes. And one of the things I would do is I would have little bits of candy and just toss them out to the little baby weasels and toss this out.

And I wish I had one. But yeah, exactly. Update your instructions file, or again, improve your context. And so one of the things that you could do is if you know for this specific type of scenario, this is the MCP server I want you to talk to, put it inside your instructions file.

And in fact, Azure, their MCP server, when you register it, VS Code will actually prompt you and go, hey, do you want to add that into your

there you go.

But if you install the Azure MCP server, it will actually say, hey, you should add this to your instructions file.

OK. That was a lot of questions.