How Software Development is Evolving with AI

Key Takeaways

[Music] Hi everyone and welcome to sip and sync with Azure. In this episode we are handling a very important crucial topic that you all have in mind. How is software engineering changing with AI and I've got Goen here who's going to show us a bunch of demos. Hey Govin, I'm very excited for this. >> Hey Priyanka, nice to be back again. >> Great. Okay, so let's start with software development life cycle. How is software engineering changing? What is your perspective? >> Yeah. Uh, thanks for this. I mean, this topic is really top of mind for, you know, almost every software engineer. Uh, there's a lot of anxiety and nervousness, but I actually have good news. I think there's going to be a lot more software engineers because I think you have to focus on where, you know, AI is bottlenecked on. And that's what we're going to talk about today. and we're going to show it uh concretely with a task that is very um I I I think a task that you really care about which is AI tours. >> Yes. >> That you've been going around and doing a great job with keynotes. >> Thank you. >> So, we're going to help you do that and we're going to assign to these agents. >> Great. Okay. So, that is a very concrete use case. I'm excited to see the first step. My my usual first step is going in and doing the research and to build the presentation and the demos. How do we begin in this case? >> Yeah. So before we kind of go concrete, I just want to kind of zoom out and give you uh the picture that we're going to like get concrete with here. So at high level like if you think about okay you know the models are going to get better uh you know this whole model forward uh approach means that as these models can plan uh reason and tool call better and better uh you can start to trust that these models can asynchronously work on tasks longunning tasks on their own right which naturally means that you can have you know five or six uh you know let's say user stories being done by you know an async agent at the same time. However, the humans uh and the software engineers are going to evolve to become tech leads. So in this graphic, what you see is you know the pre and the post of what the agents uh do uh are going to be bottlenecked and that's where you know software engineers are going to spend a lot more time and for example in this graphic what you see is first you need to be able to understand you need to be able to synthesize and decompose the user's need and work backwards. And for that you know again you can employ agents for example researcher agent from N365 copilot uh allows you to synthesize from different as different aspects of the work and create rich PRDS product requirements docs. >> Uh Devon's deep wiki uh which is one of our partners allows you to understand a codebase in a you know rich and a systems thinking way. Um we have a GitHub copilot spaces that allows you to do the same thing and then you assign it to you know these army of asynchronous agents right there's uh GitHub copilot coding agent there's codeex uh there's devon there's replet there's a bunch of them on Azure uh and then synchronously the developer can continue to work and these agents you configure them with MCP tools as well so they can uh work on longunning tasks use the tools available uh to you know kind of reflect on the tools uh response etc and keep going back and forth until the acceptance criteria is met. Now once the new codebase is there there's also day2 aspect of your workload right u and and there also we have agents like s sur agent that autonomously kind of monitor and make sure your um you know SLAs and SLOs's are met the huge a huge caveat here is you as a software engineer are underwriting the risk that the changes that you're introducing to you know application whatever or any digital experience um is on you. uh so it's very important to validate and verify what these agents are building. So what I've heard to summarize is you first need to understand the task from what whatever the requirement is uh really be able to take all compress all of that data in which we are calling context engineering these days get all the data together that is needed for the task to be done and then go into actual development which could be an army of agents then go into um testing and other things right um and then you go into deployment which could also be agents and then u monitoring out beyond deployment which could be s sur agents and other things like that. So just imagine every single software development life cycle um step would could be one or multiple agents doing different tasks. Did I get that right? >> Yeah, absolutely 100% correct. And now let's get concrete. Right. So you have this AI tour keynote uh you know site or whatever you want to call it or agents that you want to create, right? So what I've done here is you know first I understood your need hopefully and I'm going to use agents like researcher agent and I'm going to ask it to create a PRD and the cool thing with researcher agent uh that's part of M365 copilot is it uses my works context and here you can kind of see that it created this rich product requirements document and I'm going to take this and break it down with acceptance criteria and assign it to agents. Now this is where you know we can help. I have uh a few sites. Asynloom.com is one of the main uh sort of sites that you can go to and we'll have the link in the description. And here you have, you know, an army of agents that you can use on Azure. And it actually like if you go to get agu.com, that's the first site here that I have a link for. Um you can essentially, you know, customize you can first kind of see all the agents that are available. Uh the synchronous agents as well as the asynchronous agents. We have get a copilot coding agent, you know, Devon, replet, and here you can there's a setup guide so you can configure it for your repo and kind of how to get started with um and then if you already have an existing repo, you can go and get uh customized instructions for that repo. So you can go to you know github.com you know your repo and change hub to agu uh agent unblock is what it me what it means. Um you can get customized instructions. For example, here I have a repo called snippy and you can kind of see how you can customize uh that or configure that repo for in this case Devon. Um now similarly um in your case for for this AI tour um you can actually take an existing template and I have another site AI foundry.app app. Uh again, you can go to a asynccloom.com and you can you'll see that as one of the cards here. And in in AI foundry template, you can essentially take a template like let's say get started with chat because I'm assuming you want like a brand ambassador. Uh >> yeah, that can ask before and after questions we're on stage. >> Yeah. So, you can take a template like this and you can configure and customize it for your specific scenario. So you just come here, you describe your scenario, um what kind of theme you may want uh it to be and then you can break down into uh tasks or you can just go one shot and assign it to one of these coding agents um and you just have to kind of configure the agents. >> Okay, let me contextualize it. So for for this keynote, let's say I have four demos that I want to make. Sure. >> Which is what the researcher came back with as well, right? the three or four demos I want to make. I go in and here in the get started with chat. I put all of my four demo information in here and then agents would fire up. I'll select one of those to to start like maybe coding agent can go ahead and start working on one of my demos or all four of my demos and then I go tweak after. >> That's right. >> That's it. >> You got it. and and you and and you can kind of see that uh you know even let's say even before we assign this let's say you want to see a a mockup of this like maybe you want to visualize the PRD right um you can go to GitHub Spark github.com/spark >> um you can kind of see that it quickly you know builds something here based on >> just talked about yeah these are the four demos I would do >> and this is code too and you know this is actual code that you can take and you can from here assign to GitHub copilot coding agent or agent >> because this has more context again context engineering more more data that we can provide it it'll do a better job at creating the code >> yeah and and and as these agents are working we also assigned it to Devon for example and as these agents are working you want to be able to align them you want to be able to audit them um you have and you want to have a delightful experience while doing this right uh in fact I do think like uh you know as software engineers we have to we're probably going to working with six or seven agents at the same time. >> Yeah. >> And we have to be even more embedded with the user needs >> and you know you may even want to go on a walk to really reflect on the new context that you want to engineer, right? >> Uh and while you're on the walk, you may want to, you know, monitor these agents. So we have a teams app here. Uh so you can kind of see here, you can either download it and upload it to your teams or you can have your admin install it in your teams. And with this teams app, you can actually um kind of monitor the status of the agents or you can assign new tasks to the agents >> when I'm on the walk. >> When you're on the walk. That's right. This is great. >> Yeah. Uh so so these background agents uh you know can also work for example as part of your GitHub actions workflow. >> Yeah. >> So we have here a GitHub actions workflow that you can essentially uh configure as part of your repo. So every time maybe a commit goes out you know you can kind of check the unit test code coverage and if it's not you know optimal you can have uh in this case codeex you know add more uh tests to improve code coverage >> or you can you know while you're working you know in GitHub copilot you can have background agents uh and unlike our competitor we actually give you flexibility to have you know multiple uh agents whether it's get a copilot coding agent or replet or devon or codeex. So you can pick the agent that you want uh to work from GitHub copilot and it can work on five or six tasks and we have instructions here for how to configure that or from N365 copilot. You can also work with these agents from there. >> Great. Okay. So what you've done here with the async um loom site is brought some of that software development life cycle agents together. So where we started from you can see all of those kind of connected and how to get started with them on this site. We will put the links for this in the description below. Um, all right. How do you want to um want to summarize all the things you just said? >> So, in summary, I think there's going to be a lot more software engineers. I think how we do uh work. It isn't just about writing code to introduce changes. Uh it is really about you know understanding the user needs and as you said context engineering and also on the other end verifying validating because you are underwriting the risk of whatever you put out there. it's in your name at the end of the day. So, make sure that's good and and and use agents, you know, throughout the SDLC life cycle. Uh, you know, including day two aspect, right? S agent and things like that. Uh, so that output poor software engineer goes way up 10x, even 100x. >> Yeah. With that, we hope you enjoyed this episode and we are looking forward to hearing what you think about software engineering with the in the age of AI. Uh, drop us in the comments below and we would love to hear your thoughts. So with that, thank you so much. We'll see you next time.