Agentic Development Is Forcing Us to Become Better Engineers with Jon Galloway
37 minAudio
In this episode, I sat down with Jon Galloway to talk about what agentic development actually looks like in practice inside Microsoft and out in the real world. We explored how AI-assisted software development is changing fast, why prompt-only workflows are maturing into more disciplined engineering practices, and why the best results come from balancing LLMs with deterministic code, scripts, tests, and automation. A big theme throughout our conversation was that AI is pushing teams to get better at the fundamentals many of us have always known we should do more consistently: clearer specs, better documentation, stronger unit tests, better reviews, and more intentional architecture. We also got into practical tooling, including GitHub Copilot, Copilot CLI, VS Code, Visual Studio, Squad-style multi-agent workflows, MCP servers, and where these tools fit for developers versus non-technical users. We zoomed out to compare today’s AI shift with earlier platform waves like cloud and mobile, and closed with grounded advice for developers, enterprises, and everyday users: start small, make it fun, work iteratively, and treat AI like a collaborator that still needs direction, guardrails, and quality checks. Jon Galloway: Jon Galloway | LinkedIn .NET: https://learn.microsoft.com/en-us/dotnet/ .NET Rocks!: https://www.dotnetrocks.com/ Microsoft: https://www.microsoft.com/ James Montemagno: https://montemagno.com/ GitHub Copilot: https://github.com/features/copilot GitHub Copilot CLI: https://github.com/features/copilot/cli Visual Studio Code: https://code.visualstudio.com/ Visual Studio: https://visualstudio.microsoft.com/ Squad: https://github.com/bradygaster/squad Brady Gaster: https://developer.microsoft.com/en-us/blog/author/bradyg/ GitHub Actions: https://github.com/features/actions Model Context Protocol (MCP): https://modelcontextprotocol.io/ Microsoft Copilot: https://copilot.microsoft.com/ Microsoft 365 Copilot: https://www.microsoft.com/en-us/microsoft-365-copilot Microsoft Foundry: https://azure.microsoft.com/en-us/products/ai-foundry/ Microsoft Agent Framework: https://learn.microsoft.com/en-us/agent-framework/overview/ SharePoint: https://support.microsoft.com/en-us/sharepoint/platform/what-is-sharepoint ChatGPT: https://chatgpt.com/ PowerToys: https://github.com/microsoft/PowerToys Godot Engine: https://godotengine.org/ Meta Quest 3: https://www.meta.com/quest/quest-3/ Raspberry Pi: https://www.raspberrypi.com/ Google Gemini: https://gemini.google.com/
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Barry LuijbregtsAnd I've been following your career for a while. Heard you're also on dot net rocks and stuff back in the day. I've been also doing software development for a couple decades. As you've also been in the industry for a long time now as well. Yeah. What have you been up to lately?
Jon GallowayYeah. So it's interesting. Our team, so I work on James Montemagno's team at Microsoft. We've moved around a bit, but we do the same thing. So a few years ago, I was mostly focused on dot net community, and then over time, that expanded to dev dev community. And dev dev was kinda all across developer division. And then as of this year, unsurprisingly, we're focused on agentic development. So,
honestly, agentic development is kind of part of everything developers do anyways nowadays. It impacts just about everything. So so, really, it's a community team, and community is a lot of different things. So sometimes like this morning, I've been helping with I help run a few of the blogs at Microsoft. I coordinate, do editorial review, basically manage the blogs.
And then, also, I help run external events. I help run internal events. I write content. This year, I've been focused a lot more than in the past on internal kind of evangelism. Okay. And Yeah. Basically helping you know, it's a big shift, this kinda agentic development world. It's a big shift for everybody, including people inside Microsoft. And that includes, you know, traditional software developers, but then it also includes
PMs, marketing, legal, like, And I'm working with a lot of people that are, say, in marketing, have never written code before, and they're writing a lot of code or they're using a lot of code. You know? Yeah. Yeah. You know? And, you know, maybe they're using, Scout or maybe they're using Gab Copilot. Semisurprisingly, a lot of people are using Copilot CLI for, like, PM or, marketing roles, things like that. And
they're they're getting a lot done, and they also wanna be more, let me see, intentional about what they're doing. You know? They wanna they're kinda like, okay. Things are kinda working, but what can I do a little better? So we we, like, last week, we ran this internal conference. We had, look the numbers, I think, about, like, 3,000 views on this internally, and I think we can do more in the future. But it it was pretty cool to see people all across the company coming together and just learning. We had, like, a who's who of all the top people.
You know? We had keynote with Scott Hansen, and then, you know, just all kind of Of course. Yeah. Real session. I learned a ton, honestly. And, I mean, this is my job, but I still learned a ton. So
Barry LuijbregtsYeah. But that's kinda what this is. Right? When you teach, you learn, and you keep learning. And, also, this stuff is so new, and it keeps evolving so quickly. Right? It does. It really does, and it it kinda helps to have a,
Jon Gallowaya perspective that learns from history. And, you know, like, we take our existing experience in with good software development practices and good, you know, like, this is what a good program should look like, and this is this is the kind of code that's okay to just do a quick one off, and I don't care what the code looks like. And for this other project, I really do care what the code looks like. And I want tests, and I want quality and a software, like, lifecycle and things like that. So you know, we take all that, but we also have to take a lot of humility in now because things are changing. And a lot of things that I think
were accurate a month ago are not anymore. You know? And so I always have to kind of take a lot of that into, both my intuition and then when I'm talking to people, hey. Here's what I think is the right way to do this, or here's what I think this feature set is. It may be different.
Barry LuijbregtsSo Yeah. It's a that's a very strange thing. So I do a lot of GitHub Copilot workshops for developers now for customers. And every time I do one, I the day before, I just open everything up again and then I take a look at where the buttons are now or which new buttons are there. And almost always, there's one clever guy or gal there and she says, well, you know what? No. It's probably more like this. And I'm like, yeah. You're you're probably right. Yeah. Yeah. New insight. Because also the question is, you know, we're just talking into a text box basically. Right? So, it's not rocket science. It is natural language.
They made it super easy for us to use this stuff. But then everybody's asking themselves like, how can I use this better? What's the magic sauce? And then there's all these things like skills and custom instructions and custom agents and MCP servers. Mhmm. But also, those are not rocket science. It's all markdown stuff. But what do you think? Are they still very important or or not so much anymore?
Jon GallowayI think they're a lot more important than we've given them credit for. And I think that's becoming more obvious now when people are caring about and being charged at the token level. And in the past, you know, we were all operating in startup mode, it's like, who cares about tokens? Just have a loop run all weekend and just burn a million tokens. You know? And I think now we're starting to think more about, like, how can we be more efficient? I've really been I've written a lot of MCPs.
I've been writing a lot of skills. I'm also, like, doing a lot of things with, oh, like, workflows and, you know, GitHub actions and Yeah. You know, really trying to think about anything that I can do deterministically, do with an EXE or a script or something. So, like, just this past, this past few days, I was writing up a report on this conference, and I was like, hey. I want a bunch of screenshots from all these
all these videos. And the videos are all in SharePoint and you know? So but I was like, hey. I could use AI and just say, please go get a bunch of screenshots. But then I thought about it. I'm like, wait. This is super inefficient, and I don't know what I'm gonna get. And it may run for a half hour and then give me trash. So instead, I had it write a script, get screenshots from five and fifteen minutes into one video. Okay. That looks good. Now do it for all the videos. You know? And and, so I'm I'm just really focusing a lot more. That's something I,
like, as a, you know, tech evangelist, really think is something I want to be, you know, speaking about and and communicating better is like, hey. Things that you the the getting the right balance between what an LLM is good at and what deterministic code is good at is really important. It's gonna be it's gonna save you money, but it's also going to give you more predictable results. Higher quality results too.
Barry LuijbregtsYeah. And and, yeah, higher quality equals less tokens.
Jon GallowayExactly. Exactly. Because you're not repeating work. You know, we we've all kind of we're going through this, this growing up phase of working with LON. And the first kind of hello world is like you talk into a chat box, and it actually gives you back useful results. And you're like, woah, this is crazy. Like, I remember I remember closing every chatbot for years because they were stupid. And then all of a sudden, one time on ChatGPT, people are like, hey, try this ChatGPT, and I did something, and it worked. I'm like, this is crazy. But then after a bit, we we think, hey. This is not super productive typing. Like, for instance, I would, way back in the day, like, type chat, write this function. No. That's not right. Do it again. No. Do it again. And and so that's that's a really, like, naive way of doing it. Right? And then we're kinda growing up over time to
I started working more with, like, do some things and do pull requests and have the pull request do a co code review. You know? And then and then over time, it's like, hey. What can we automate in terms of those GitHub actions? And, you know, just working with this whole kind of you know, really kinda seeing it as more of, like, an outsource or, like, you're working with another you know, someone else is writing code. That's fine. We've done this and other you know, this isn't completely new to collaborate with another developer. No. Now let's put quality into the process. Like, I want unit tests. I want, you know, like, I want,
like, rubber duck style reviews. Different people review each other's code, etcetera. So
Barry LuijbregtsYeah. So it's it's basically like you're outsourcing it to some some country, your software development. Right? Exactly. You're not sure what you're gonna get back, and you don't know about the quality. Not really. Right. And I've done this at a lot of other jobs in the past, and sometimes
Jon Gallowaythe problem is my fault because I didn't give clear clear specs. Right? You know? And so, like, it really comes down to instead of saying build me a website, it's, hey. I want you know, I'll actually start now with a PRD, a product requirements document. And I'll I'll say, here's you know, I'll work with AI to write a a document in markdown with very clear, like, here's what I want. Here are some references. Here's some similar existing projects, websites, open source software, or whatever.
Here are the acceptance criteria. Here's the, you know, initial, phase delivery and then future possible things. And and so really hammer that out and then say, okay. Iterate on phase one. Right? And and the quality is so much better because it doesn't go off and build me something in PHP because I never told it. Right? So Okay. You know, if I say actually say, no. You know, I want it in this language. I want with this testing format. I want this, you know, color scheme, whatever. Right? So
Barry LuijbregtsYeah. Exact and and that is already spec driven development kind of. Right? Because you make a spec, and then you you do the thing. And we've we should have always been doing this. Exactly. And that's part of the thing is that, you know, AI is good at
Jon Gallowaypredictable grunt work. So things that are frustrating and boring, and that's things like, you know, writing up specs and writing unit tests and, you know, writing documentation and doing architectural reviews for duplicate code. And, you know, all these things that are Yeah. Not hard work. They're just boring, and we all hated it. So we didn't do it enough. You know? And so now it's a time where we can just put that in the process and start integrating that.
Barry LuijbregtsYeah. Yeah. So I see that a lot in companies as well, where they have to basically go back into their processes and mature themselves with more specifications, more unit tests, more quality controls around what they're doing. Because now they hire these outsourced developers who work somewhere in a data center in The US somewhere, and you don't know what you're going to get back. Usually it's amazing, but you need some more control. We all need you know, learn to grow up. So and we're all learning because it's super early.
Jon GallowayYou know? And on on the one hand, like, part part of that sounds really boring. Right? But then the the nice thing is that allows me to focus on the fun part, which is the creative part. So, like, on, you know, weekends and evenings, I've been building I've I've built a game for my, my Quest three VR headset. I've built a iPad, stationary bike app. I've built, you know, games with Godot. I've been building, you know, just, like, websites, all kinds of fun stuff.
And I get to focus on, like, what is the game script? Is this game fun? How can I, you know, add a zombie mode to my biking app? Like, whatever. You know? And and that's the fun part. And I have tons of unit tests, and I have tons of specs, and I have tons of quality steps
Barry Luijbregtsso that I don't have to think about that stuff. You know? Yeah. And what's your favorite, let's say, modality to use, GitHub Copilot in this case? I
Jon Gallowaybounce around a lot, and partly because that's my job. I need to be able to Mhmm. Keep up with everything. Lately, I have been using a lot of Copilot CLI and Versus Code, and partly that's because of just, like, the different projects I'm working in. But I do, of course, use I use the the new GitHub app. I've been really liking the canvases in there, which are pretty fun. I made a canvas for Squad. I don't I don't know if you've looked at Squad, but I made, like, a dashboard for Squad, and it keeps track of all the different things.
And then, of course, Visual Studio. I use Visual Studio a lot for things like modernization. Yeah. I was really good at that stuff. But, I'd have to say a lot of the time, I'm just using Copilot CLI with Squad. I for my personal projects, I've been really liking using Squad and, like, dialing in the, the different Squad members and, you know, having, like, ones that are very focused with all the skills. You know, the front end designer has,
say, like, specific skills, and the and the tester has and so then they can be very efficient, use just the right model. And and it's really fun to just kinda set it all up and then say, like, implement phase three, and then go have a copy and come back. Yeah.
Barry LuijbregtsYeah. For for listeners to who don't know what SQUAD is, it is kind of this open source framework by Brady Gaster.
Jon GallowayMhmm. He still owns it, I think. He does. He co run runs it. There's, I think, Tamir, and there's several other people that contribute to it. Mhmm. And it's based on I think there's some other kind of similar things and other AI kind of there's, I think, Gastown and some things like that. The general idea is you have you have agents in it working together in a team, like collaborative team, and you have a coordinator, and then you have specific roles. So you'll have, like, a front end dev and a back end dev and a and an architect and a tester.
And, you know, I've seen some kinda people say, oh, that sounds silly. And, you know, I I kinda get it, and I kinda like, once I work with it enough and got it tuned in just right for my projects, I get it. You know? Like, it really and the the important thing is that LLMs are constrained by context. Right? There's only and so over time, as you keep chatting back and forth, you're building up all this context. And then in order to like,
when that context gets full, then we have to compact it. And so then we lose some of the information because we've compressed down that that chat. And so over time, we're kind of always battling this problem of, like, so much context going back and forth in the conversation. And what's really cool with these sub agents is that we can really focus the context. So I can have issues assigned to a sub agent, and the sub agent is is not concerned with any of the other thing. You know? Like, if the front end agent doesn't worry about other things, they can ask the other agent or it can get a summary from something. But it
and and also, it can be very tuned as far as what the skills are, what MCPs it works with, what language model it's using. And so because of that, like, for instance, if I have a scribe role or a Git, you know, a role that's just focused on my Git history and things like that, it can be a very fast model. It doesn't need to be super smart. I don't need to waste tokens, and I can also use a very tuned model, you know, that's very quick. So so that's one of those things, again, where just like in the real world, you wouldn't just hire one super genius to build your project for. You you'd probably hire a team, and the team would have,
you know, peep and and also, you can tell I'm getting all excited about this. But it is fun. You can have, like, the the agents can learn over time, and you can say, like, hey. I didn't like what you did with all the purple gradients there. This is the style I'm going for. I want you to remember this. And that specific agent will make that note. So, you know, it sounds like I'm selling Brady's project, and I do like it. But it's really more of a pattern. And, you know, it's just built using markdown and skills
and, you know, like, it's it's just a pattern that you can use anywhere. GitHub actions and things like that. But the idea is like, hey. If we orchestrate things so that we've got you know, these are front end skills. These are back end skills. These are architectural skills, testing skills, etcetera. Then you can split that off, and you can you can be more hyper focused with that. Yeah.
Barry LuijbregtsYeah. Yeah. It's very cool, and, it it's been out for quite a while. So I really am wondering if this will ever get absorbed into the product as a feature or or some sort of template or something like that or a spin off product. I don't know. Yeah. I know that the teams talk together a lot, and there are some specific
Jon Gallowayfeatures in in Copilot for, like, fleet mode and some of these other specific they're they enable the general ideas, but they're not and so I know that they're kinda figuring out because they also squad is kinda prescriptive and may not fit in specifically with some people. So so at a product level, I think they're trying to get that sort of balance right. But I think generally across, you know, not just the copilot world, but the AI development world, that's a
a thing that it's it's not the right thing for every project, but it's the right thing for a lot of projects. Yeah.
Barry LuijbregtsYeah. Definitely. Yeah. It's interesting. So we're evolving so incredibly fast. Like yeah. If we zoom out now, right? So we have this whole thing that now developers have a new tool basically to be more productive. Companies as well where they can stick AI into applications with, I don't know, with Foundry, Microsoft Agent Framework, things like that to create a chatbot or something more interesting perhaps. We see all this happening also with customers that we
talk to from our side. It's very interesting. But if we zoom back out, right, this is one major technology evolution, revolution. We've seen more. Like, back in the day, we had cloud, for instance, was kind of a new I
Jon Gallowaylove that you brought that up. I'm sorry to jump in, but this is something I've been thinking too lately is people act like this is out of like, we've never had this happen before. We've never gone through this. And the cloud revolution was huge. You know? And there were a lot of different ones over time, you know, even going 16 to 32 to 64 bit and, like Yeah. You know, open source. And, I mean, we've gone through a lot of big software revolutions. Outsourcing was a big thing for a while where it was like, oh, no. This is gonna change everything. Yeah. And,
you know, even the Internet, it kinda the cloud wave went in some big phases and, you know, like things like, oh, I don't know, the the JavaScript revolution, you know, where JavaScript all of a sudden was running everything. We're like, how did that happen? So it it is it is important to, like, take it seriously and to really kind of, like, not ignore it and just say, like, I I don't do AI. Like, it's that would be the same as saying, like, I don't believe the Internet exists or, you know, like Yeah. I I think you have to recognize that it's there, but also,
I don't know. Put it in context and think about, like, hey. We've gone through change. The whole software development industry has been a period of rapid change, you know, over the past couple decades.
Barry LuijbregtsSo so we we had big ones. Right? Like clouds. Big one. Mhmm. We still have, know, you still companies that are still hybrid or moving to the cloud, for instance, for all sorts of reasons. We had smaller ones like big data, smaller ones. Oh, yeah. Remember? The big data. Mobile first was also one of the things. That's right. That's right. Yeah. The metaverse kind of thing that never really, really did anything. Yeah. And now there's this thing. So do
you think this is different profoundly, than, let's say, cloud? Because that's kind of a big one.
Jon GallowayThat is a tough question. I do think it is profoundly different, but I think some of it is that this is this has more of an opportunity to accelerate. Cloud can so for instance, a cloud was an architectural change, and while speed kinda continued to go up and we have more computers network together, the cloud you know, well, some of the patterns and things have changed over the past
decade or so, but it's not like a fundamentally different cloud now than it was ten years ago, say. Whereas, large language models, I do feel like we're still we've been growing in terms of what the models do. I I kinda feel like we're on this edge of, like, some sort of architectural shift and, where it's it's likely to keep accelerating more. And then I I do wonder about things like,
you know, quantum as quantum gets thrown into the mix with with AI. Right? And I think there are some huge opportunities for for you you know, like, when the math nerds all of a sudden are like, hey. We found a way to make things twice as fast or twice as whatever. So I I feel like we're, in a way, kinda scratching the surface of what large language model technology is.
Barry LuijbregtsYeah.
Jon GallowaySo I don't I mean, that that's kind of gut driven. That's kind of a hunch sort of feel. But I do feel like this is a bigger change partly because software development is all about code. It's implemented in code and everything, and large language models are extremely good at writing code. So it's the whole loop of I'm stuck. Go look it up on Google and Stack Overflow. Find some code. Paste it in. It didn't quite work. Okay. Do this. It's instant now. Right? And
I was thinking about the this past you know, like, a week ago, I ran into something that was frustrating me in PowerToys, which is in c plus plus and or the the thing was and I'm like, I'll just fix it. Right? And I, like, iterated. And, of course, I was very careful. I haven't written production c plus plus in a long, long time. But I feel like I actually, like, got this thing done in a short amount of time relatively. And, I I guess where I'm going with this is that that I feel like we're going to be changing the building blocks of software development
using large language models in a way that that accelerates more, if that makes sense. Yeah. Yeah. I'm still figuring this out. I don't know the answer, but it just feels a little bit of a bigger shift than some of those other ones were. Some of those two were a a little bit style driven. Like, hey. It's mobile first. Okay. That's a consumer preference. Right? You know? And it's like, hey. I would like things to work better on my phone. Whereas large language models open up
whole new things where, like, people are who aren't even develop my younger brother is not a trained software developer, and he's building something with Python run running on a Raspberry Pi that it's like, this could be a product you could sell. You know? I mean, it's it so it's kind of really a big shift.
Barry LuijbregtsYeah. Yeah. I I think so too. That it's very different. Because if you look like, at cloud, for instance, that is very deterministic. These are products that we now sell. You just move your stuff. Your your computer is just now over there instead of over there. Yeah. And we package it nicely so that you can use it easily. And that kept evolving, of course, but still it's deterministic. You just do it or you don't. Yeah. And now we're in this strange world where we talk to things that predict predict the next words, and it's nondeterministic.
And I think as humans, we have a bit of trouble adjusting to this this world of nondeterministic technology that kinda does stuff, and it can do a lot. It's yeah. It is wild. And and I think intuitively,
Jon GallowayI kept thinking of it. It took me a while to realize it's just text in and text out. There's no, like, okay. It acts like it's text in and text out, but it's really doing some other things. It's like, no. It's really just predicting the next token. You know? And Yeah. We keep we keep adding layers on top of it with better harnesses and better system prompts and better, you know, corrections and tuning. But at the end of the day, it's just a big predictive model of what's the next token.
Barry LuijbregtsYeah. So you mentioned that you're you're teaching this also to non techies Mhmm. Within the company as well as perhaps also outside of the company. Yeah. Yeah. So where's kind of the the line of, you know, when do they use, let's say, Office three Copilot three sixty five, which is the business version, versus, GitHub Copilot? Because you can kinda use both, right, for almost anything.
Jon GallowayYeah. That's also a slightly tricky question because, like you said, we have multiple GitHub Copilot, some things to go. Right? And I feel like what what I try and I try and look at what is the most comfortable, easy fit for the thing you're doing right now. Right? So, like That's good. My family, you know, I'll show them Copilot on the desktop or the web Copilot app. You know? And then, you know, that's a quick, like, I can get started, and I can use this basically like a better Google, you know, kind of Yeah. Okay. So I feel like that's a nice kind of
if you're doing office work, I look at m three sixty five Copilot. You know? And then with more sophisticated users like Scout. And that's because it's connecting to, you know, company resources and and that sort of stuff. And then I do feel like if you're doing anything code related, it really helps to be in an environment that understands code base. So, like, that's where, you know, GitHub Copilot is really productive because it
can figure out the Git history. It has my entire context of files in the solution or project or folder. Right? And, like so that was something where I mentioned my younger brother was building something with Python, and he was doing everything in a really long chat GPT conversation. And that kinda works, but it kinda doesn't after a while because it loses the context, and it's, you know, it's doesn't really so I I kinda steered him over to working in Versus Code. And and,
you know, it's like, look. Here now you have all your files in one folder, and now I've got a cleaner context. Right? So I feel like that's but it really is kind of a where do you feel most comfortable? And, you know, we we got that question internally at that conference. It was like, hey. You have all these different you know, do I use Copilot CLI or the GitHub app, or do I use Versus Code? And it's like, well, get get familiar with the different ones, and then use the one that feels best at the time. And I very often will start a project.
A lot of the time my flow is, I'll just pop into the Copilot, the web app, and I'll say, hey. I'm thinking about this sort of project. Is there something similar already? You know, how would it work? What would and I'll iterate for a bit, and then I'll think, I think I actually wanna build this. I'll say, write me a markdown block that I can paste in as a PRD. And then I'll pop into, like, Copilot CLI. And depending if I'm feeling like I'm ready to you know, where I'm at, I'll I'll initialize a squad, and I'll say,
here's a PRD. Start this project. And then, you know, let get things going. And then after a bit, I'll say, I really wanna dig into this code more. And then I'll pop into, you know, Versus Code or Visual Studio and, like, get the whole view of it. So and then, you know, so I really feel like it's nice to like, you're not stuck in any of them, especially when you're developing code. Right? You can bounce around. You can and so a lot of time, I'll have, say, Versus Code and Copilot CLI open, and I'll work with both of them. Because it's all the code's,
you know, in the same directory. It's I've got a git, you know, keeping me safe as long as I'm committing regularly and you know? So yeah. Yeah. Yeah. Don't have I don't have a super clear like, I can talk about, well, if you if you wanna look if you wanna work at the chat level and you wanna manage multiple sessions and and things like that, you might work with GitHub app more. And if you wanna, you know, work more if you're in a space where command line is kinda feeling simpler, you know, then Copilot's got and if you have an existing code pace, you might wanna pop open Versus Code or Visual Studio and kinda,
you know, make it easier to see your kinda code context. And so but, really, I feel like it's just like whenever I feel yeah. I'm a little constrained here. I'm a little claustrophobic. Right? I wanna break out. I wanna move to the world that you know? So so and and GitHub Copilot's with you in any of those places is, you know, like, marketing thing, I think. So it's like the same models, the same context. As long as you're preserving things in markdown docs or issues or whatever, it just goes with you anyways.
Barry LuijbregtsSo Yeah. Yeah. And and even sessions sync now to GitHub. So per perhaps they can also sync to different clients in a bit.
Jon GallowayAnd, you know, another thing too there is, well, it feels like this large language model is conscious, and it's sitting there in this chat with you. It really is kinda like a stateless web where every single chat question that goes in, it's a brand new thing. Right? And so it, in a way, it doesn't really matter. As long as I'm preserving my context, my issues, my session state like you're talking about, then my next chat, even if I switch harnesses and go from CLI to Versus Code, it's the same thing. Know? Hey.
Now I wanna fix this issue. Okay. Go ahead. You know? And Yeah. Yeah. I I think that's where it's going. Definitely. Mhmm. Yeah. Mhmm.
Barry LuijbregtsOkay. As we we're nearing the end here, you've already given lots of advice and talking about also non techies and stuff. So what would you advise somebody that's kinda in tech or maybe not, and and has heard of AI but is kinda still on the fence like, nah. You know what? I'll sit this one out. Let's see what the next thing is. What would you say to them?
Jon GallowayWell, I would I guess, like, I've I've recommended in a lot of things over time is, like, give it a try in something small. Give it a try in something fun. For instance, my younger daughter, you know, has written a a game for her college project. I said, hey. Let's sit at the kitchen counter, and let's work on building a game, or let's build a website together. It is really helpful if you do have a nerdy friend who's already using stuff effectively.
Just sit down together and solve a problem together. You know? I think it is don't don't focus on the like, it's very easy to get caught up in the who's winning now in this race, who's who's the biggest player, who's the you know? Like, none of that matters. What what works for you when you're solving a project? I do recommend using AI modes if it's, you know, if it's, Copilot consumer or if it's Google Gemini or whatever. Try using that instead of Google searches
and use it for all kinds of things. I use it for cooking recipes. I use it for figuring out, you know, home repair problems. Like, just all sorts of things like that. Right? Yeah. And so so I I feel like that is a good a good first step for consumers. It's just instead of Google searches and kinda trial and error, use those those sorts of things. And then for developers, I recommend trying trying a project but thinking about it like you would in real professional software development. Meaning, don't say build me a website, and then it builds a bad website, and then you say AI is terrible. Think about like you communicate with a junior developer.
Right? And write up a a detailed plan and then iterate through it. And the same way that you would kind of build trust with a a new developer where you wouldn't say, go build an entire project. I'll check when you're done. Say, like, implement the first thing. Implement this. And, you know, like, and then over time, as you build up that velocity, then you say, okay. Now spec out phase two. Implement phase two, you know, and kind of step back as as you go.
And I do one final thing is for enterprises, for people wanting to use this in production environments, focus on quality things that you haven't been doing, like unit tests, like documentation, like architectural reviews, like, you know, performance analysis, benchmarks, things like that that are not gonna break your application, but are likely to actually improve your quality.
And once you get familiar with those and comfortable with those, that also makes it easier to implement AI because you've got all those safety nets of, you know, your unit tests and, you know, all those even documentation. AI can read your documentation to build better code. Right? So if you have good quality docs, that improves your velocity there too.
Barry LuijbregtsOkay. That that's great. Yeah. That's great stuff. Okay. Thank you very much. So where can people go and connect with you?
Jon GallowayThat's a trickier thing. Social media has got a little crazy. I'm on Blue Sky mostly. I've been thinking about getting more involved in LinkedIn kinda in the future. But and then, you know, also at conferences. Whenever I'm at conferences, love to chat with people, come up, and say hi. So
Barry LuijbregtsOkay. Real real world preference. Yeah? Yep. That's good. Just just go and say hi. That's we will do that. There you go. Excellent. Alright. And thank you very much for this conversation.
In this episode, I sat down with Albert Tanure, a Senior Cloud Solution Architect at Microsoft, to talk about what AI is really changing in day-to-day engineering work - and what still matters just as much as ever. We...
In this episode, I sat down with Heena Refai, a Cloud Solution Architect at Microsoft, to talk about how enterprise AI has evolved from the big-data and traditional machine learning era into today’s generative AI and...
In this episode, I sat down with Rob Bos to talk about what’s happening across GitHub, Azure DevOps, GitHub Advanced Security, and the fast-moving world of GitHub Copilot. Rob shared how his work evolved from DevOps...
In this episode, I sat down with Anton Sizikov from GitHub to talk about how AI coding tools have evolved from “smart autocomplete” into full-blown agentic workflows. We unpacked the early days of GitHub Copilot, when...
In this episode, I sat down with Pascal van der Heiden to talk about how fast AI-assisted development is changing the day-to-day reality of building software. We got into his current workflow across Azure, GitHub...