The Evolving Role of Developers in Tech: AI, Infrastructure & What’s Next | Agents Of Dev Ep. 3

Control. This is agent dev. I'm in position.

Copy that. Dev. Stand by for, go.

Standing by. Hi everybody. Welcome to Agents of Dev podcast.

I'm Mitch Ashley, and I lead the software lifecycle engineering practice at Futurum. Also a practitioner in my history. Of course, we're talking about all things, Deb.

And I'm my co-host Brad Shiman, who is also analyst and practitioners. Hey, hello, Brad. Introduce yourself.

Hello everyone. Uh, Brad Shiman. I lead our practice on data intelligence, analytics and infrastructure here at Futurum.

And, uh, Mitch, it's really nice to, to chat with you again today. And, um, I understand we're, we're gonna be talking about, uh, my favorite subject, which is the, the battle for the hearts and minds of, of developers. Yeah.

Uh, so what happened? There's a great disturbance in the force. I thought we didn't need all these developers anymore.

Right? You're like, they're gonna have to go find other jobs and AI was gonna do it all, probably like for every other job, right? The over over rotate on the hype around ai and suddenly, suddenly folks got religion about, Hey, developers are important again, matter of fact, we better go get those developers because that's who's gonna build on our platform, use our technology, use our frontier models.

If we're, if we're that, and I'm, and I'm speaking about, you know, what happened at AWS reinvent, they, they kicked off like, developers, we love you. Here's our IDE, here's the things that we're doing. You know, I lot, I loved it.

I stood and said to the ceo, I'm like, this is great. I never expected this to happen, but, you know, God bless you. I didn't say that, but God bless you for loving the developers again.

You know, we bless you. Yes. And and the other thing, you know, Google, Google is trying to sort of corral its cats of mini doors into the Google developer universe, um, which is a different challenge.

Um, of course, Microsoft. Yeah. There, there's a lot of spaghetti on that fridge.

Yeah. There are some of have been on their own, right? Oh, yes, exactly.

And, and we're not, we're not scraping it off. We love it. Yeah, exactly.

It's, it's history. It's art now. So, you know, I have my thoughts on, on why, what's your perspective on like, what, what happened?

What, you know, what bit got flipped? Some dip switch in a board somewhere in That's that's, that's a pun. Or you, you, what is that a four year slip?

I dunno, now that I said it, it's my goodness. Yes, it does. You know, that a wanted a abandon that, that we seem to, to have, you know, embarked on a, a year or so ago with, uh, yeah.

AI is just gonna replace everybody. And any job that involves code and the written word will be gone in a year's time, um, has certainly changed. Or at least our perspective, uh, on it has changed.

And you need to look no further than our own, our own beloved, um, capital here in the United States, which just announced a new program to hire back, uh, the folks who actually make this stuff run. Um, it's, it's, you know, it's, it's a testament to just how important infrastructure is. And I think when we look at a few of the outages we've seen and and why those outages happens, uh, it all comes down to that infrastructure and keeping the lights on and how critical that is.

And so you kind of couple that realization, that wake up call together with the, the fact that as you and I have chatted about before, um, companies like Microsoft learned many, many years ago, that if you want to capitalize on the enterprise, you need to win the hearts and minds of developers yeahs minds sheet. Yep. Yeah, yeah.

You gotta have the, you gotta have the tooling. The tooling is what developers gravitate. And I'm using that word, you know, with all its many different meanings toward mm-hmm.

Because you know that gravity has impacts on who signs for what on the dotted line. It's, you know, what's, what's fascinating too is sort of in parallel, not not unrelated, very, very related is I've heard, um, Google and AWS say, we're not an infrastructure cloud, we're an application cloud. So repositioning even, well, at least for that audience anyway, they're, they're an infrastructure cloud.

No doubt. They definitely are. But it, it's interesting to hear them just say that this is coming out of the developer, kind of, but e even at, at AWS reinvent, I think that came from the CEO if I remember right?

Well, yeah. Reflected in the tooling they rolled out. Huh.

You know, it was, let's build apps. Let's not build infrastructure. But, you know, you and I both know that when you build the apps, they run on infrastructure.

Uh, so I think it's the bringing me your tired, your poor, and your AI applications to run on my infrastructure. Yes. All of them.

And there are, or there are a lot. Right? Right.

Don't you do anything, Mitch. There's like, uh, this sudden embarrassment of riches right now for, for tooling. And, and if you look at like the Google ecosystem, uh, my goodness, you know, I was actually just yesterday chatting with one of our, uh, folks internally who, you know, um, asked me, she said, so I see that in notebook, lm I can create a slide presentation from the resources that I upload, you know, PDFs and whatnot.

She says, can you, could you put that into, uh, this competitive intelligence tool that we're working on as a, as a project? And I'm, and I said, I would love to, um, but I'd have to build it from scratch. 'cause it's not in their API yet, and I use the word yet because I guarantee that that will be in the API because Google is aggressively, and I think we're gonna chat about one of those, um, a little later.

Google is aggressively building out its API to incorporate increasingly higher levels of abstraction and, and more specific types of functionality that you can build on top of Gemini and all of the, uh, vertex AI resources. Well, this is, this is in part at least anyway, the whole, is it all AI agents or is it all structured languages? You with API calls and procedural?

Well, no, it's never all one or the other. It's usually a hybrid and pretty big hybrid, I think in this case. Agreed.

So, so she brought up the Google API question. I think it was Monday of this week, uh, may have been last, again, it was Monday this week. Google announced this interactions API for its deep research model, um, which is not just an API, it's an API to essentially an agent that's acting on your behalf, um, doing work for you, but it brings with it, um, structures, brings with it kind of memory and managing what it's doing over a long run.

Um, and it also brings, um, more of a way of observing what's happening in your interactions opposed to that's critical. Random sets of, of prompts that I might generate or I might use or whatever. It, it kind of puts more like programming what we're used to into the picture.

So it's like structure meets, you know, uh, dynamic output. It's interesting. Yeah.

The rigor is important, isn't it? And observability is important. And when I saw that, uh, my first thought was, oh, thank God.

Now I don't have to use Lang Chain and Lang Graph. And please, no offense to the folks making that software because it is fantastic old Langs die old Langs graph. Well, it's, but you know, it ev everyone knows that, um, you know, if you're building something to scale in the enterprise, that lang change you portfolio, while capable is more difficult because it isn't fully documented, it does change.

Um, and it's a little harder to integrate and to manage. And so when I say thank God, I'm like, yeah, you know, I as a builder am a nerd in the, um, you know, Google, uh, vertex AI ecosystem. And so seeing that build on, as you just mentioned, building on this new deep research, API call that they just released a couple weeks back mm-hmm.

Says to me, wow, I I can use a consistent A-P-I-I-I have just one set of documentation that I need to manage and, and look at when I'm building something. And I like that, you know, it, and it's interesting that they did it with deep research for long running, kind of more involved type of interactions. So maybe you don't need it for kind of smaller interactions, or if you do, let's work on, let's, let's introduce this in a more challenging scenario where you really are gonna need this a p that interaction.

So it things that, that they, I, I get picked up of it is about this sort of idea of fragile prompt chaining where you have ad hoc state handling happening when you're just doing prompts between, in processes, between talking with agents, and of course, you know, procedural structured languages don't think in prompts, right. They think no. Well, unless it's, it's just a variable.

Unless it's my code, then, then it thinks randomly, but no, I'm just, oh, does it it does it. Oh, nice. Yeah.

I like that. It's a mind of own that's using Rand. Yes.

Turn off mind of my own brandand. Um, but so, so it was interesting to phrase it that way, um, in thinking about, you know, the sustained chain of prompts. Yeah.

And that being a fragile interaction. I hadn't really thought about it that way. Yeah.

It's an interesting transition. And, uh, again, we saw that reflected in reinvent at reinvents with AWS talking about, uh, long running processes for code to, you know, code, uh, let's say like a code review to not just like add a feature, but instead to, um, optimize, you know, let's say you have a huge repo repository, multiple repositories all working together, and you're looking at your bill going, Ooh, you know, what happened there? Well, you know, the tools that, that AWS rolled out with its, um, hero based, uh, agent tooling.

Yeah. The frontier model, I think is the, the frontier models they called them, right? Yes.

The long running, but it's all, it's all on Nova. Um, and, and, and Claude, mostly Claude, I think, um, yeah, there's a lot of Claude, but, uh, at any rate, it's, the idea behind it was to have a process that maybe kicks off on Monday and doesn't finish until Friday, and, you know, when it's met it's objective and it's objective may be a high level objective, not a, you know, add a button here, but instead a optimize my spend. Maybe that's the prompt.

Mm-hmm. I, I don't know. It could, it could, you know, work that way because aren't these thinking reasoning, um, routines within those large language models, not very good at things like, uh, ascertaining the, uh, you know, the intent of a que a query based upon the context.

True. Yeah. Ex Well, interesting.

And, and sort of procedural language isn't gonna be great about context either. Right? You know what, what, what if makes me think of too is we're, we're in this stage of transitioning from everything as a prompt interaction like we do with, with ides and, you know, with the clients, et cetera.

More than chatbots, just our interactions with models to, um, something that's more systemic or systematic. And 'cause today, you know, when you're doing development with an ai IDE, you're managing sessions in session state across, because you can't only, you can only go so far in, in a session, then you've gotta start another one up. And, you know, everything I've read and I've started to do is okay, just work on one change in one session, then start up another one.

That's right. Document it Yeah. And document it.

Yeah, exactly. Because you get drift and suddenly, you know, what was working well, all of A doesn't work well. Oh, yeah.

And I was working with, um, chat GPT five two, or excuse me, GPT five two over the weekend just doing a little project. And I was like, dang, that that worked. Well, th this is really an improvement in what they've done in the sort of constraining the model from trying to write the world's most largest set of code for you that you could ever ask for that you didn't.

Yeah. Right. They all have a different personality with how they write code.

That's very true. But they all like to write code. I mean, could I write that for you?

Please, please. It's kind of the big at your, yeah. I think like Claude is the, the, the most verbose, I would say, of those that I've worked with.

You know, it's good code. Mm-hmm. But my goodness, do you really need to go that far?

Yep, yep. Well, it's, it's, you know, we're going into 2026 and we're working on our predictions, and I'm thinking about things like, okay, so how is the, how is the interaction model changing of how we create software both in the tooling, but also in the infrastructure of the software, whether it's agent control plane, now we're seeing more APIs, so I think we'll see a lot more APIs like this that gives you better structured. And I think that actually opens AI up to more people to use it, who are sort of, I, I don't know about this un kind of dynamic, you know, I might get a different answer every time.

Gimme a way to interface with this. It gives it a lot more people that can use it. Yeah.

You see that, don't you? In in how the, these companies that we work with every day are building out their portfolios. And look at Google's, um, you've got their traditional Gemini chat, which can create, uh, artifacts.

You've got, uh, collab, which is that spaghetti on the fridge that will never scrape off because I adore that tooling. Um, you, you've got newish things. Um, I don't even know.

We, we, um, you remember Project I dx the Google, I do remember idx. Yeah. It didn't go away.

It's now Firebase Studio. Yeah, exactly. Is reincarnated into fire based tune.

Yeah, yeah. Which knows that we'll reincarnate into, right, right. Yeah.

And then you have like the, the pure no-code, you know, ideas like Opal, uh, which, which is fascinating. And, and it looks like, you know, n innate n uh, without all the BS attached to it. Um, and then you have, you know, the more developer ish, but still very user friendly Google AI studio.

And then you have, um, you know, notebook lm, which I mentioned earlier, which, you know, is, is very much a user, uh, experience, but, you know, how far away is that from simply being a part of the infrastructure that developers in a, in a company can work with? I would say not very far. I gotta believe it's can happen.

I mean, it's, it's, it's already sort of part of Google Workspace, but not quite yet. You know, it's, it's a standalone thing, but that's gonna be part of it. Like, here, I want these documents, I want this stuff to be in my notebooks, use this folder as a notebook, blah, blah, blah.

Yeah. Now I'll put this as a full stack app and, and hosted on Firebase with Firebase. Ooh.

There's some dot connection going on there. Boy, you're reading the head in the book, aren't you? You're on chapter.

It's just fascinating. Yeah, it really is. Well, we probably should, uh, you know, go to our next segment, so it is time for the drop.

Okay. It's time for the drop. Okay.

We're back. I love our drama and our traffic stuff. I'm so entertaining.

You did such a great job with that, you and Andy and everybody. Um, so I'll, I'll, I'll jump on the bandwagon. You know, it sort of what's on my brain this week is there's a lot of things to wrap up by end of the year.

So we're working on predictions. We are, um, tuning our practice areas. And someone actually was an AR person said to me yesterday, said, uh, well, you know, uh, the, those predictions that you're working on, don't get too comfortable with them because you'll wanna change 'em in about six months.

I'm like, you, you are very right. So my time horizon of what is the prediction has changed from, you know, in 2030, this is what's gonna happen. If you're high enough level, you can do that.

And you really think about the work that people are doing with AI that changes monthly, almost, almost weekly. So I'm, I'm trying to reset what I think a prediction is both some long term and also short term. How about you?

What's the drop for you? 5. Right.

And, and maybe that's more in line with how fast the world seems to be evolving right now. Maybe we, maybe an annual, you know, budget annual, this annual that is a little too slow for us. Um, but, but seriously, uh, right, right now what's on my mind is, uh, you know, uh, cultivating a, a bit of disdain for MCP, um, and, uh, and also for, um, for training cutoff dates.

Uh, and, and, uh, the fact that as I've, as I've mentioned, I think in one of our earlier podcasts, it's, it's been very frustrating, uh, wherein you, you have an a set of documents, an API that you're working with, and as we just talked about, hey, there's a new API call, but we wanna do, and that one might be deep research, and I wanted to implement it, uh, last week. And so I, I said, you know, working within Gemini CLI, Hey, let's do this man, here's this new API and I, you know, had it do a URL context, which is a tool that Google has because, you know, they own the index of everything on the web, so they don't even have to scrape websites anymore. You just give it the URL and it has all the data.

Um, and it's still still reverted to the API calls from 2024, um, in setting it up. It was very frustrating. And I thought maybe what, what would be best to do is, is simply to do it more programmatically instead of, oh, and by the way, I should mention, I even had it consult context seven to, to look it up.

Okay. And it failed, um, with that and, um, right. And, um, so I, I said maybe, or thought maybe what I should do, and I haven't tried this yet, but to do it more programmatically through an API call, because, you know, you can query a function and, and see what it does.

Mm-hmm. Mm-hmm. You know, if I import the Gemini, um, you know, library for, uh, their models, I can query them and say, how do you know, what are you expecting from this function call?

What are the parameters? What are the outputs? And maybe that's how, you know, I should be thinking about adding functionality is to, you know, start with that baseline of what the import is for the library that I'm using, and then just inquiry after the library that's loaded, that's, that's running in the app.

I mean, you're not gonna go back to the man pages and figure out what does this API do? Well, apparently it ignores that. It doesn't care so much.

Not so much for that. No, you, you bring up a really great point, which is, you know, when, when technology transform is truly transformative, how you interact and how you use it changes. Yeah.

And it sometimes breaks the model, our mental model, like you're talking about it, well, maybe it isn't looking it up in the code library and trying to understand what it does and walk through whatever I need to see what the parameters are, and maybe there is some documentation, maybe not now it's code. Explain now it's query, query explain. Right, right.

Tell me what this does. How, how do I best call this? What are the kind of results, if I call it this way versus that, you know, kind of interaction.

Yeah. It's just a tool call at that point. It is like, you know, every thinking model in particular, but all of the agent tech tools that we have have built in LSPs or, um, you know, pals for, for doing just like Python execution to do basic math and stuff.

And so why can't it just, you know, use the code that it's got. Mm-hmm. 5, we couldn't do that.

Yes. Fractions too. Couldn't, because we would have to start January would've to be zero.

So that Yeah. Which is before you, we started over and named, came up with our calendaring system. That's right.

I, I shift all my arrays, so they start at one. You do. Well, good.

Good man. I, I'll be sure to remember that if I look at your code. Oh crap.

That's why it's that. Boy, that would really mess with people for sure. Alright, Brad, it's been fun.

Um, we won't see each other on the road for a while, but we will see, hear each other here on Agents of Dev. So thanks for doing this podcast together. This is a lot of fun.

Yeah, I'm enjoying it too. And thank you everyone for tuning in. We're, we're looking forward to seeing you next time.

We are. com. Love to hear from you, feedback, comment on the, on the, uh, podcast platform that you particularly use.

And like we'd love to hear from you. We, we will absolutely watch comments and hear the, the, uh, the criticisms and the, and the praise or anything in between. See what you like to hear.

We are having guests come on, so we're kind of tuning up, scheduling up who we think we might have on. I think one of our first folks will be the CTO from FUT who's building our intelligence platform. And he's looking forward to that.

He's a pretty brilliant guy. He is done some amazing things, so I'm excited to talk with him. Well, we'll announce that coming up here, but we have other people interested in being on, so if you're, if you're interested in talking about agents of Dev, uh, reach out to us at that email address.

We'd love to hear from you. So, well, Brad, signing off. Have a good rest of the week.

We'll see everybody back on the next episode. Thanks for listening and watching. Thanks everybody.

See you. Mitch. Control.

This is agent dev. I'm in position. Copy that Dev standing by.