“[Software is] gonna make a lot more API calls. It's gonna do it in new ways with new combinations. I think the handwritten software was always kind of limited. You can only ship so many versions of your app every year.”
The rise of AI agents is more than a tooling upgrade - it's a fundamental rewiring of the entire developer experience, with your APIs at the very center.
We're joined by Matt DeBergalis, co-founder and then-CTO-now-CEO (congrats Matt!) of Apollo GraphQL, to explore this massive transformation. He introduces the emerging concept of "agent experience," explaining why systems built for human developers are not ready for the unprecedented scale of AI calling APIs.
Matt argues that as the old rules of software development get re-evaluated, engineering leaders must rethink their entire stack. He presents a powerful analogy: a structured data layer like a graph is the perfect "left brain" for the "right brain" creativity of LLMs. This provides the semantic precision and guardrails needed for AI to act reliably, enabling a future where user experiences are personalized "to 11" and APIs become the core business asset. This conversation is a crucial guide for leaders on how to prepare by prioritizing higher-level system design, and why clear communication and architecture are becoming far more critical than handwriting code.
Show Notes
- Explore Apollo GraphQL's graph infrastructure and MCP tooling: ApolloDev
- Connect with Matt on LinkedIn
- Connect with Andrew on LinkedIn
Transcript
(Disclaimer: may contain unintentionally confusing, inaccurate and/or amusing transcription errors)
[00:00:00] Andrew Zigler: Hey, everyone who's been listening today, we're joined by someone who has helped define how modern software gets built in many different ways.
[00:00:07] Andrew Zigler: His name's Matt DeBergalis, the co-founder and CTO of Apollo, graph ql, and the co-creator of Meteor Js. And if you've been listening to the pod lately, you know, we've been talking about how AI is colliding with traditional APIs and the teams that build and manage them. And the rise of agents means a lot of different things for software development.
[00:00:27] Andrew Zigler: But this shift isn't just a tooling upgrade, it's a rewiring of the entire developer experience. And today we're going deep on what those things like agent experience really mean, and how Graph is at the center of it all. Matt, welcome to the show.
[00:00:42] Matt DeBergalis: Thanks for having me. It's good to be here.
[00:00:44] Andrew Zigler: Great. Well, let's go ahead and dive in 'cause we have some cool stuff to talk about today.
[00:00:47] Andrew Zigler: You know, we've been talking a bit about this new term agent experience. It's kind of like a complimentary to developer experience and it, it seems like it may be an evolution. Some people see it as a replacement. There's lots of people who have different opinions on [00:01:00] this, but Matt, what do you think about this emerging concept of agent experience?
[00:01:05] Andrew Zigler: Is it really any different?
[00:01:07] Matt DeBergalis: it's such an exciting time and I've been. Building software for a few decades now, and I've never seen anything like this in terms of just how quickly things are changing. So I suspect we have a lot more to learn. But the cool thing about bringing AI into the development experience. Is just what it lets us rethink, across different parts of the stack. And you know, I do this 'cause I want to see lots of great software in the world. I think we need more of it. I think we need it coming from people with more backgrounds and different perspectives. so for those of us in the dev tool space who think about infrastructure and the plumbing that makes all this stuff go under the hood. the rise of AI is like this new superpower that We can use in different ways and we can talk about what this means for, for creating APIs or calling APIs or lots of different parts of the stack. But I think it's gonna change a [00:02:00] lot, right? I think ais don't work the way that people work.
[00:02:04] Matt DeBergalis: And, some of the things that we're secondary concerns when we were building. for people to use are gonna come to the fore. some of the things that maybe felt really important to get right, turn out to be less critical, and I just try to come in with an open mind. 'cause I don't think anybody has the answer.
[00:02:23] Matt DeBergalis: And, and I think only, exploring the stuff is gonna get us there.
[00:02:27] Andrew Zigler: Yeah, I completely agree about, there's a lot of software to be built now and there's a lot of scrambling to figure out how we build it. Um, I think it's a really insightful call out that you say that some things move more to the forefront in terms of what we need to be concerned about.
[00:02:40] Andrew Zigler: Some things maybe take more of a backseat because now they're simpler or easier to approach. So it's really like a reshuffling maybe of the priorities and what's important.
[00:02:48] Matt DeBergalis: that's just my fancy caveat for whatever I say next may not be true.
[00:02:52] Andrew Zigler: Right.
[00:02:53] Matt DeBergalis: I, you know, we've certainly learned a lot about, what does have to change and. I just think it's important to, [00:03:00] to remember we're on, day one of big transformation in how all this works.
[00:03:04] Andrew Zigler: And this transformation, do you think it starts at any particular point of the developer experience or API orchestration? Like is there like an obvious starting spot or is it kind of attacking it from all sides?
[00:03:16] Matt DeBergalis: I mean, there was a, there was a chapter of AI right away where the LLMs were clearly helpful assistance for the engineers and really everybody in software development. And obviously they've gotten better, but even in the early days, I think we all had these experiences that made us rethink, Hey, I've, I've got the ultimate pair programmer sitting next to me, or, you know, I'm a tech lead.
[00:03:39] Matt DeBergalis: Now I'm, I'm not just gonna gonna write the software myself. there's another chapter that's just emerged, uh, anthropic announces MCP. A few months back, and to us, it's sort of the starting gun for now. We can build agents. Right now there's a, a, an accepted way that Theis can actually talk to external systems. that [00:04:00] leads to a whole other category of stuff that we can think about doing. I think that one's a lot more nascent, but it's, it's near and dear to my heart because Apollo is, at the end of the day, we're an API company. We're about that lets you connect your software to your APIs and agent that can call APIs as an agent.
[00:04:17] Matt DeBergalis: That's actually useful. It can do stuff, right? If you think about. All the interesting actions you might want to take, if you wanna put something in your shopping cart or get a shipping price estimate or recommend a product for another user. You know, these are all, the end of the day, gonna be API calls. It's just a lot more interesting than the early chapter of user facing AI that was about, you know, training the model or maybe a, a database that could augment the context with, with Rag. I don't think we've really, all kind of come to terms with how quickly this may go and how, deep that rabbit hole is. If you just think about all the interactions [00:05:00] that people have with businesses and software and how much of that might now be in reach for a, a good AI that's connected to the rest of the business, that's gonna be interesting. And we're trying to. Ask, like what can we do to help that, help shape, that help make, not just get there faster but more safely, in, in a way that's like built for the long haul and, you know, there's a lot of exciting opportunity there.
[00:05:22] Andrew Zigler: Yeah, I think there's a lot of opportunity, for people to experiment, but like you said, to also figure out what's the paradigm that's going to survive, what's the pattern of implementation that's going to really prevail here? and as people experiment and, you know, we're building agents, we're giving agents tools, it sounds like we're entering a world where API consumption is going to go up, maybe go up a lot.
[00:05:43] Andrew Zigler: Do you, do you feel that way? Yeah.
[00:05:46] Matt DeBergalis: Yeah. And, and. we're, we're the GraphQL company. Um, just
[00:05:50] Andrew Zigler: Right.
[00:05:51] Matt DeBergalis: the stage here, we, we are, excited about GraphQL because it's a query language for your APIs. what it means is in a world of [00:06:00] microservices and cloud systems and SaaS, maybe 10, 20 years ago, you only had a handful of APIs that you built your software on.
[00:06:06] Matt DeBergalis: But now you've got a typical app, maybe sits on 50 or a hundred different APIs, and you need infrastructure, you need machinery. To help orchestrate how all those API calls get made. The idea that you're gonna hand write code doesn't really scale when you start talking about, systems of this size.
[00:06:25] Andrew Zigler: Right.
[00:06:25] Matt DeBergalis: And so we've been working for years now on how to drive more value from APIs that people have. One of the things, a lot of GraphQL users, First come to GraphQL for is they've got an API that was built for one use case, but now there's this desire to ship a different kind of user experience. you're building a mobile application that drove a lot of the early excitement around GraphQL in our customers, there's this push toward meeting The end user more where they are. Like if you think about a modern hotel experience, there's probably a kiosk in the lobby where you can check in.
[00:06:59] Matt DeBergalis: That's just another set of [00:07:00] API calls. It's a different kind of software.
[00:07:02] Andrew Zigler: Right.
[00:07:02] Matt DeBergalis: you can unlock your room from your phone. Maybe when you go in your room, your name's on the tv and there's a welcome message and maybe you can, ask for a late checkout during your reservation flow online. Uh.
[00:07:14] Andrew Zigler: Right.
[00:07:15] Matt DeBergalis: And, and that's an example.
[00:07:16] Matt DeBergalis: All of those are systems that are being repurposed and, you know, kind of recombined or composed into this new kind of experience. I think AI is just gonna be like the mother of that, right? Because it's
[00:07:27] Andrew Zigler: Hmm.
[00:07:29] Matt DeBergalis: And if you imagine it's gonna be like in a year or two talking to an AI or using an AI first application. It's gonna make a lot more API calls. It's gonna do it in new ways with new combinations. I think the handwritten software was always kind of limited. You can only ship so many versions of your app every year.
[00:07:50] Andrew Zigler: Right.
[00:07:50] Matt DeBergalis: have that limit anymore. So we may also find that the pace of all this stuff just hits a new level that we're gonna have to get used to.
[00:07:59] Matt DeBergalis: and that has [00:08:00] implications too for how quickly this stuff gets added.
[00:08:02] Andrew Zigler: Yeah, and definitely a compounding speed factor. It seems to go faster and faster the more sophisticated it gets. And, you know, your position and what you've been doing at GraphQL and working on graph is. It puts you actually in this really unique spot of expertise because you already know that, you can't just throw an API on top of everything.
[00:08:21] Andrew Zigler: You need to have a structured, orchestrated way of accessing your data that is flexible for all the different experiences you want to provide. And then now that's evolving, like you said, into an MCP type experience. And you see some companies that see this and just try to map like one-to-one API to MCP.
[00:08:38] Andrew Zigler: People talk about how that doesn't necessarily work. And it reminds me a lot of how what GraphQL does on top of an API, because it, it allows you to ask for specifically what you want. it's a structured query language, and that's semantics is what is so important to it. and I think that that has a lot of parallels with MCP.
[00:08:57] Andrew Zigler: do you feel the same or do you think that it's kind of evolving in [00:09:00] a different way?
[00:09:00] Matt DeBergalis: That, that's a good example of where, agent experience is gonna be a little different from the human experience or the developer experience. it's really clear from work we've done and what we've done together with customers that the semantic precision of. GraphQL is a great fit for the models. The models are really good about reasoning and GraphQL. You know, if you haven't used it, it's strongly typed.
[00:09:29] Andrew Zigler: Right.
[00:09:30] Matt DeBergalis: It's self-documenting, so you can put English language doc strings, and that's been a best practice in the GraphQL world for a long time. There's even more than that. there's a place to put comments, and we've found sort of along the lines of, you know, an LLMs text file.
[00:09:42] Matt DeBergalis: You can, talk to the model in the definition of the graph in a pretty, helpful way. It turns out to help the model navigate. And at the end of the day, the graph is a set of objects that mean something. the big insight of GraphQL is instead of looking at your APIs as a set of [00:10:00] separate endpoints that just return j you can look at all those APIs together as a network of objects
[00:10:07] Andrew Zigler: Right.
[00:10:07] Matt DeBergalis: and the, the properties of each object.
[00:10:09] Matt DeBergalis: Maybe that's a, a product object or a review object, or a shopping cart object. If you think about an e-commerce, example, the properties of those objects are provided by different APIs and they're scattered all over the place. That's just the nature of what happens when you build software, bringing that all back home and giving the model. A precise construction of, the meaning of all this stuff has profound impact, on what it can do and what can it understand.
[00:10:36] Andrew Zigler: That structured and queryable intent behind the data, unlocks better outcomes for AI systems as people experiment. 'cause it's just like what you mentioned about you go to a hotel and you have all of these different unique experiences. That are all probably their own, APIs somewhere managed by some team, but they all are very different from each other, but they work together to create a cohesive experience of, I'm at the [00:11:00] hotel, this is convenient, I can get what I want.
[00:11:01] Andrew Zigler: Right. And it sounds like what you're hinting at is that, the layer that we're building now with agents, it has that same amount of. Personalization and customization, right? It's like, you, you really want to be able to take all of your systems, all of your APIs, all of your tools and products, and put it into one spot where the LLM can decide with its own reasoning and logic for what it wants, what to pick and choose and call.
[00:11:25] Andrew Zigler: And then that, again, going back to what you said is why. Graph is so powerful for this because it's gonna allow it to access exactly what it wants. And this is actually, it sounds like going to unlock scaling. because going back to what we mentioned, like API, consumption's gonna go through the roof. We talked about this recently with a guest on the show, um, Sagar Batchu of Speakeasy.
[00:11:46] Andrew Zigler: and. As I talk with more leaders, it, it seems to feel the same from them as well. And so if, if you're building tooling and you're trying to orchestrate this, you know, what does that look like in the future that's different [00:12:00] from today? I.
[00:12:01] Matt DeBergalis: I think this is a great example where engineering leaders. Need to rethink some of the, conventional wisdom of the last generation because of ai. Like take microservices as an example. So there's been this era of development where we said, okay, the way you scale an organization, the way you organize your teams you build individual services that can be called independently and composed together, and it goes back to the Jeff Bezos, memo at Amazon, right? Everything's gotta be an API and there's no other interface into your system. It's a great model and of what made that model work is that it, created this really clear responsibility.
[00:12:42] Matt DeBergalis: Your job as a service owner is, is a very clear, tightly scoped responsibility. We could talk about SLAs, we could talk about performance, metrics and somebody else, some human or some other team is gonna make use of that alongside a bunch of [00:13:00] other APIs to create a piece of software. AI changes that a bit because you can't just rely on the AI to do all that work that humans did.
[00:13:09] Matt DeBergalis: that's pretty high level architectural work, right? I mean, AI can write some code that'll call an API for you, but asking it to build a whole system that correctly and, and always the same way, combines 10, 20, 30, 50 API calls and handles all of the stuff around like errors and retry and all the subtlety that goes with that.
[00:13:30] Andrew Zigler: Right.
[00:13:30] Matt DeBergalis: bit much. So you're an API owner, if you're a team leader for an API, one of the things AI is gonna do is it's gonna force you to think more broadly about how is this all gonna fit together and what do I need to do with my API to make this, part of a, a working system when a lot of the software is being built with the assistance of ai. I think that's gonna force people to take more of a [00:14:00] platform view of the world because we're gonna need infrastructure at all different levels of the stack. AI is very, very good when you pair it with a semantic, you know, declarative piece of infrastructure. lots of people are using AI Help write react components. For example, there's another declarative architecture, but I don't think anybody's seriously talking about dumping react and just asking an LLM to fit all the components together. Right? that's the job of the infrastructure. The pressure to combine stuff in new ways and the fact that the AI's pretty good at parts of that and pretty turned around with other parts of that, I think forces a lot of us that are tech leaders to really zoom out and view the whole picture.
[00:14:44] Andrew Zigler: the nature of LLMs being probabilistic, right? Going back to what you just said about, you know, you can't necessarily rely on them to make all of the calls correctly in the right way, handle all the errors, because there can be lots of deviations and when you're in the world of APIs, like you don't wanna deviate from what [00:15:00] the is giving you.
[00:15:01] Andrew Zigler: And so, yeah.
[00:15:03] Matt DeBergalis: and even, and even how they combine, like I, I, uh, we shipped a an MCP server today for Apollo. this allows you to use AI to talk to your graph. And, I wrote a blog post and I talk about this. You know, imagine a, a bank. You don't want one of your customers if you're, if you're asking the AI agent for your recent transactions, you don't want one of your customers to get back five and the next one to get back eight.
[00:15:25] Matt DeBergalis: and you don't want, you know, sometimes it includes like the remaining balance along with the transaction and sometimes it doesn't. You know what I mean? Like, it's, it's more than just how to combine the APIs. it's all these decisions that we used to make in the code. About exactly what we want, how we want it presented exactly what combinations, who's allowed to see what that, is so important because at the end of the day, this all comes back to like, we're making software for people and we want people to have a great experience.
[00:15:51] Matt DeBergalis: And it's, it's really important to hit that note, I think. I don't think there's a single Apollo user that I've talked to that isn't. [00:16:00] Building an agent isn't thinking about what these experiences are gonna be like, but the flip side of it is all that stuff is still on the way to production. We're really early in actually rolling this stuff out.
[00:16:11] Matt DeBergalis: the reason is because getting this stuff right is part of what has to happen if
[00:16:16] Andrew Zigler: Right.
[00:16:17] Matt DeBergalis: if you're serious about, putting that kind of software in front of people.
[00:16:20] Andrew Zigler: Yeah. I mean, if you're gonna ship these kinds of experiences that are so sophisticated with LLM you need to have insight into how they work. You need to be able to reliably understand their outcomes. And it, it goes back to like, LLMs, they, they're, they're pairing well. With strong semantic like infrastructure and orchestration, things that help keep them on the rails, like the LLM can be the creative kind of composer of all of these different tools and, and pieces, but it needs to be guided by a rigid system that tells it yes no, and keeps it in line.
[00:16:54] Andrew Zigler: And so it sounds like that pairs really, really nicely with, you know, semantic data [00:17:00] that you get from, from a graph. It feels like a natural ingredient within that mix.
[00:17:04] Matt DeBergalis: Yeah, it's a left brain, right brain thing,
[00:17:07] Andrew Zigler: Yeah.
[00:17:07] Matt DeBergalis: they work very well together.
[00:17:09] Andrew Zigler: Yeah. And so as these things get more sophisticated and we have more, ability to do nuance, get nuanced information and make nuanced requests in a chat or otherwise using a, a tool with natural language, it kind of starts to open up a world where you get a. You can kind of solve everything with a conversation.
[00:17:29] Andrew Zigler: Do you think that we might evolve into a, place where most applications and ways of going about doing things on a computer, reside within a kind of like a chat type of interface? Or do you think that chat is limiting and that there's maybe in a additional way that this is all going to be evolving?
[00:17:46] Matt DeBergalis: I think we're gonna find out. When mobile came out, I mean, in some sense, a mobile app is just a computer app on a smaller screen, right? But it turns out multi-touch was important and some of the user [00:18:00] interface things that work really well on a desktop. work at all on mobile and, and vice versa. Right. I think mobile gave rise to a world of map-based experiences.
[00:18:10] Matt DeBergalis: For example, it's not like we didn't have maps on the desktop, but it's very different
[00:18:13] Andrew Zigler: Right,
[00:18:14] Matt DeBergalis: hand. I
[00:18:15] Andrew Zigler: right.
[00:18:15] Matt DeBergalis: play out with ai. Like we know some of the answers. We know there's gonna be this, this natural language back and forth. I suspect a lot of that's gonna be spoken. but we also know. these generative systems are really good at building graphics. So I think we might find some of this stuff turns the application inside out. We'll still have UI elements and a lot of what's on the screen today, but it might be all sort of surrounded by a much more free form interface. think one of the really neat things about AI is that it's so flexible and, and it's like personalization turned up to 11, right?
[00:18:56] Andrew Zigler: Yeah.
[00:18:56] Matt DeBergalis: Everybody. if you go to your, banking [00:19:00] application today, you and me get the exact same app. That's just how it has to be. maybe it turns out you prefer a different way of seeing your accounts and all the things you can do there than I do. And I wouldn't be surprised at all if get to the point where The typical banking experience is one that learns from your preferences and adapts to you. But underneath it, to your point a minute ago, there's a, a structure and a consistency. So we're not, we're not saying different things when we talk about the meaning of a wire transfer or a, a, a deposited to your bank account.
[00:19:35] Matt DeBergalis: I, I think we all want that stuff to be, common. But there's so much room on top of that to adapt to customers and, and to what they want. And I, I think that's gonna attract people. I think people want that kind of an experience when they're using software,
[00:19:49] Andrew Zigler: Yeah, it's just gonna like, be wearing different clothes, like just the fit, specifically what you are looking for. We already see some of this in applications where, sure, everyone gets the same app, but maybe when you land on that app, [00:20:00] you get like a, a different set of personalized topics or features or assets or, you know, you customize your home on, on the app.
[00:20:07] Andrew Zigler: Right? So it's like there's already a world where, users are invited. To customize their software and kind of like collaborate with it. And now we're gonna turn it up to 11, like you said. I, I really like that as a, as a way of like describing it. 'cause it's like just going a little, it's going beyond even the bounds of what we thought that this stuff could do.
[00:20:25] Andrew Zigler: Uh, and your prediction too. it reminds me of one from past guests. He's been on the show a few times. Rob Zuber of CircleCI, he talked about how.
[00:20:33] Matt DeBergalis: Yeah.
[00:20:33] Andrew Zigler: yeah. Yeah. He, made a prediction at, uh, an event that Dev Interrupted had hosted, back in December about how maybe in the future, when you make a website request, or you try to query that data instead of getting a website and then the website's data is hydrated for you on demand, the website itself could be created on the fly to fit what you're looking for.
[00:20:52] Andrew Zigler: So do you feel the same way? Like, that might be a future.
[00:20:54] Matt DeBergalis: something like that. I'm, I'm sure we've got all the exact details wrong. I mean, it's, it's, it's telling [00:21:00] the future is so. But personalization is, is definitely part of it. I think AI is probably the next big, push to elevate, not just the, the, the amount of API calls that you were getting at before, but just think about how important APIs become because APIs are the capabilities of a business.
[00:21:19] Matt DeBergalis: Like it's the stuff that a, a company or a system can do.
[00:21:23] Andrew Zigler: Yeah.
[00:21:24] Matt DeBergalis: you know, Uber, we all use the Uber app, but maybe the app gets a lot less important, and there's already hints of this, right? you can call an Uber from Apple Maps, or maybe it's Lyft or I, I don't know. There's all kinds of like
[00:21:36] Andrew Zigler: Right.
[00:21:37] Matt DeBergalis: and, and and map partnerships
[00:21:38] Andrew Zigler: Okay.
[00:21:39] Matt DeBergalis: As an example. Uh, but the real value of Uber in a real sense is the, is the network of drivers. the machinery that knows how to schedule, knows how to price. that's the actual asset, I think. So it's companies that have something like that and can [00:22:00] expose it in a agent friendly format. That's just a fancy way of saying good APIs.
[00:22:05] Matt DeBergalis: I think, that might have profound impact on, companies that we think of as, Really valuable or, or, or that have a, a, a really important experience. And I think there's gonna be a race in a lot of different industries to go in that direction. You probably won't be able to use your user facing application surface area as your calling card in the future.
[00:22:30] Andrew Zigler: Right. It becomes, you know, having that killer app that has like all of the good experience in it, it's still valuable, but it's maybe not like the clincher that it would have been in the past because now it's about the asset of, you know, what is the service or the data that you're providing as part of your company.
[00:22:46] Andrew Zigler: People want to consume. With Uber, it's like you said, it's the scheduling, it's the availability of the drivers. It'sthe networking of all of those through their APIs. And so in a world, there's like a world where, you would call your Uber from like a conversation or another [00:23:00] app entirely, and so you get almost like a commoditization of the services that people are providing.
[00:23:06] Andrew Zigler: And so. if you're an engineering leader right now, like a lot of our listeners are, and maybe you're, maybe you already have one of these very valuable types of data sets or APIs or orchestration that you're sitting on top of, that you, that you ship and put out through an app or a website or a SaaS, whatever, whatever it may be, what are some steps that you think that that leader should be taking today to make sure that that asset is unlocked tomorrow?
[00:23:33] Matt DeBergalis: Well, I think they should go to Apollo dev and, uh, I mean, I,
[00:23:39] Andrew Zigler: it's all, it all starts with a graph for sure. Yeah.
[00:23:42] Matt DeBergalis: I do think platforms are gonna be really important. I think every engineering leader should have a really clear platform strategy in terms of what's my stack,
[00:23:55] Andrew Zigler: I.
[00:23:55] Matt DeBergalis: components of that stack?
[00:23:57] Matt DeBergalis: and in the areas where I'm [00:24:00] maybe relying on bespoke software that we build ourselves, maybe look for a, a standards based platform based. Alternative to that. I feel pretty good about that prediction because know that AI generated software constraint, I mean, a good platform basically provides guardrails and rules and laws of physics, right?
[00:24:22] Matt DeBergalis: So when you,
[00:24:22] Andrew Zigler: Right.
[00:24:23] Matt DeBergalis: that, we know that the AI can go very, very fast. My guess is that aren't gonna be writing much of that software in just a few years. Right? It'll just be the AI and. Maybe there'll be another AI that does your code review and a third AI that does your, you know, CI/CD sequencing.
[00:24:40] Matt DeBergalis: And I'm sure there's a human somewhere in that picture, but it's gonna start to get pretty abstract. But I think there's gonna be this growing need for a, a higher level understanding of the overall system and the architecture. that's exciting. I mean, I, I, I think like. We don't write assembly code anymore, right?
[00:24:59] Matt DeBergalis: Like we, this [00:25:00] whole
[00:25:00] Andrew Zigler: Right.
[00:25:01] Matt DeBergalis: is just a series of increasingly powerful abstractions. To use them, well, you have to understand the layer underneath that abstraction, right? You, you weren't a good C programmer if you didn't understand assembly, but you didn't have to write assembly anymore.
[00:25:14] Andrew Zigler: Right.
[00:25:15] Matt DeBergalis: you could say the same thing about other modern programming languages.
[00:25:18] Matt DeBergalis: I think the same thing's gonna be true about. How all this stuff gets glued together. You'll have to understand Kubernetes, but you probably won't have to write operators too much anymore. You'll have to understand how graph works. I think that will be the orchestration layer for APIs, but you probably won't be writing GraphQL schemas by hand or queries by hand. have to understand how React works. But I think the days of handwriting react components are probably not much longer for us. Right. That's, that's my guess. And I think you can work backwards from that. You know, we find an Apollo just internally, we've really leaned into the culture around writing, and documentation because I think it's a good way to do that kind of higher level thinking [00:26:00] and design work that you need to do. I just think there's a lot of interesting implications for how you build an engineering team, what skills you emphasize, what skills are maybe a lot less, less critical than they were before.
[00:26:12] Andrew Zigler: So strong communication skills, strong understanding of the engineering fundamentals. Those are the things that are, are, are more important. And doing things like actually writing the code takes more of a backseat because of, of what you described this future where LLMs are gonna be doing a lot of that process.
[00:26:29] Matt DeBergalis: Even thinking about it from the point of view of how you talk to a if, if you're using Cursor or
[00:26:35] Andrew Zigler: Yeah.
[00:26:36] Matt DeBergalis: any of these things today, right. What I find is I. If I know what I want and I explain it clearly, I get good results.
[00:26:43] Andrew Zigler: Yeah.
[00:26:43] Matt DeBergalis: those are the two skills, right? You have to know what you want, which means you have to have a certain, just higher level knowledge of, of the broader landscape, and you have to be able to explain it well and it's just interesting.
[00:26:55] Matt DeBergalis: These just, they're different skills than, you know, certainly. you know, we [00:27:00] emphasized a generation ago That shouldn't be a big surprise to us. Right. The, the, the technologies change and, the skillset changes with it, but that's the kind of thing we're thinking a lot about. And, it's exciting because I, I think it opens the door for a lot of really strong, talent to do this kind of work that,
[00:27:17] Andrew Zigler: Mm-hmm.
[00:27:18] Matt DeBergalis: didn't have some of the more. Domain specific knowledge of, of React or, you know, whatever else it may be. And, and, and didn't feel like they could build the software today.
[00:27:28] Andrew Zigler: Yeah, I really identify with that myself. 'cause I have a humanities background. Before I moved into Dev Rel, which was kind of like my bridge into text. I used to be a teacher. Um, I studied like classical languages, right? And so it doesn't have really have anything to do with programming. Uh, but as I've gotten more, you know, I've learned to code.
[00:27:44] Andrew Zigler: I was in web dev for a while and as I've, I've moved more into like developer marketing now I see this world where. These communication skills that I spent so long building and using it to understand the underlying concepts of something and communicate what I want. Like those are becoming even more important.
[00:27:59] Andrew Zigler: [00:28:00] And these are soft skills that engineers have, you know, maybe in the past not prioritized as much, but now become so much more important. and it opens the door for a lot of people. To then develop their own software. We've talked about this in the past about, you know, the, you'll get the rise of like personalized software, disposable software that you just make once to do something or for yourself or for a project.
[00:28:22] Andrew Zigler: and you know, we're seeing this really fast right now on, the web and with web dev we talked with. Uh, we talked with Vercel Lee Robinson over there about this because, you know,
[00:28:30] Matt DeBergalis: Yeah.
[00:28:30] Andrew Zigler: moves, nothing moves faster than the web. And so they're right now kind of seeing this where people make stuff on demand and people from all sorts of non-technical backgrounds are able to get in there and make something.
[00:28:42] Andrew Zigler: But at the end of the day, they have to know what they want to make. And so having that higher level understanding of what is my goal in going on Cursor and asking it to build something, it kind of contrasts the whole, thing going on right now where everyone's talking about vibe coding and everyone has a different reaction to it.[00:29:00]
[00:29:00] Andrew Zigler: Some people have a very negative reaction to it, and some people think that it's like an accessibility thing for many people to be able to get into programming. But at the core of all of those conversations, everyone always arrives at this same point of, you need to know what you want and you need to know when the LLM got it right or wrong.
[00:29:18] Andrew Zigler: You need to have that, uh, that, that understanding. What's, what's your take on that right now?
[00:29:22] Matt DeBergalis: Yeah, I, I agree. I think the other big gap, and it's just an age old story, is there's a huge difference between prototype and production. And
[00:29:31] Andrew Zigler: Yep.
[00:29:32] Matt DeBergalis: I think that was true long before Vibe coding, right? Like lots of things can get built in a sprint that for real, take the better part of a year and AI's gonna accelerate all of those constant, I don't think it's gonna change the fundamentals where. the software we actually work with every day as people, the stuff on our phone or the stuff that we use when we, when we log in for the morning, so much more to it than [00:30:00] like, what was on the screen and, and getting a first version of that to work.
[00:30:03] Andrew Zigler: Yeah,
[00:30:04] Matt DeBergalis: And I think sometimes that doesn't get the attention that,
[00:30:06] Andrew Zigler: I.
[00:30:06] Matt DeBergalis: the, you know, the flashy stuff gets, but it's no less critical and. I'm sure we'll see. This seems like a pretty safe prediction. I'm sure we'll see a lot of going into that part of the development lifecycle, in terms of where AI can help and what ought to change about this stack. But I know for sure that, the typical team that we talk to has a mandate to ship.
[00:30:32] Matt DeBergalis: They are, they're being asked to prioritize ai. they've got some instincts for what the experience ought to be, but there's a lot of open questions, not just the ai, API orchestration problem, but there's a lot of questions about what the rest of that stack's gonna look like. you're into dev tools or infrastructure, it's a very exciting time because all this stuff is, is gonna get sorted I think, over the next year or two. And, the thing I hear over and [00:31:00] over again is you can't just YOLO and AI experience and, and not worry about the part where it, you know, leak sensitive information or, just think of all the things that could go wrong, right? And, and people are well aware of that, so. there's still a lot of work ahead of us, I think, for all of us to get that, to get that stuff squared away.
[00:31:21] Andrew Zigler: Yeah. we're hinting at this world where. The ais are, writing the code, they're testing it, they're reviewing the prs, and, and, and now they're also at the heart of accessing your data, understanding the structure of your data. You know, what, what kind of prevents this from becoming the wild west?
[00:31:37] Andrew Zigler: How do we really keep a handle on it as engineers, and understand what's going on when all of these systems are interacting.
[00:31:45] Matt DeBergalis: Yeah. I mean, I fall back on system design for a lot of this, right? If you, if you think about the overall stack, the LMS just one part of it, and if you think about The responsibility, the role of each piece of the stack. And if you think about the sort of inherent limitations or safeguards [00:32:00] that, the boundaries between those things create, that's part of the answer, right?
[00:32:03] Matt DeBergalis: So like I may not understand exactly what the L LMS doing, but I know that if I give it a very specific set of MCP tools and I understand what those tools do, that creates some amount of safety, right? I don't think the AI is gonna. Reach around MCP in the near future and get what it wants directly out of the database.
[00:32:23] Andrew Zigler: Right.
[00:32:23] Matt DeBergalis: may, maybe there's a, a, a dystopia that's coming for us here, but I, I think, I think the data's safe for the moment. you know, it's stuff like that that, that I think like covers the bases around a lot of this. I think analytics and measurement is another huge unanswered question, right? What, what's the right way to measure the quality of a ag agentic experience?
[00:32:49] Matt DeBergalis: It's gonna be a lot different from how we think about CSAT and software. And, my guess is that a lot of that will get fed back through some kind of an offline AI process. Like, one of the things we're thinking about [00:33:00] is you don't wanna give AI unfettered access to all your APIs. That that's just a non-starter. so the way that we approach this and, and what we talk about today with, with our MCP tooling is that you use the graph infrastructure to create a set of specific MCP tools, right? Maybe there's a tool for adding something to your shopping cart under the hood that might call five or 10 different APIs.
[00:33:22] Matt DeBergalis: But the way those are combined is, is done on the graph layer. So the AI's not making something up here, but the question becomes, how do we know what tools to build? And how do we know that when we put a customer in front of the agent, they're getting a good experience? 'cause you've, you've probably had the experience where you called a human agent for something and they were like, I can't really help you. Okay.
[00:33:43] Andrew Zigler: Yeah.
[00:33:44] Matt DeBergalis: So I think there's gonna be a whole flow around feeding the whole interaction through a different set of processes with a human in the loop. Maybe this is what a modern product management job looks like, or even a. Development [00:34:00] job. It's hard to say, right? I think there's gonna be some new job titles here, there's a, there's a process where we learn from the interactions in real time. Turn that right around into sipping a new capability for a customer. it's, it's interesting. This is what I was getting at before, like if I opened my app on my phone. Say my app for my bank, and it was different every day. I don't think anybody really wants that,
[00:34:26] Matt DeBergalis: but, but agents can get better every day and it's just good. It just means they're more capable. It means they make fewer mistakes. It means I. have to fall back to the handwritten software or maybe an interaction that involves a human less often, which means I get it done faster, right? Like that's just net good. And so I think that we've, we've removed some of the natural bottleneck on how fast this stuff can improve. And that's where I think there's gonna be a really rich vein of, AI enabled products and technology that help us do that well.[00:35:00]
[00:35:01] Andrew Zigler: I completely agree. I think there's so much opportunity that we're right on the cusp of, of taking advantage of, and you've kind of outlined in this conversation a lot of really great ways that teams can get started with it and some really salient predictions I think about where all of this is going.
[00:35:15] Andrew Zigler: And for me, you know, this has been like a really fun chat, Matt, you know, you've such good insight on, on what's evolving and the industry right now. But before we start wrapping up, where can our audience go to learn more about Apollo and the work y'all are doing? You, you mentioned you recently dropped an MCP server.
[00:35:30] Andrew Zigler: I.
[00:35:31] Matt DeBergalis: Yeah, we're Apollo Dev, so you can read about the underlying graph infrastructure as well as, what it means to that with an agent and the MCP stuff is cool. I gotta tell you, like, if, if you haven't had the experience yet, you know, one of the reasons GraphQL flourished is that it's, it's a joy. It's, it's just, it's really fun to
[00:35:54] Andrew Zigler: Agree.
[00:35:56] Matt DeBergalis: It's, tools are solid, but there's something even deeper than that. It's [00:36:00] just like being able to, to explore your APIs in this way and, and navigate that stuff. I, I think any developer that's had that experience knows what I'm talking about. There's a new one now, which is watching an LLM crawl over your graph and write a query.
[00:36:14] Matt DeBergalis: It's just, it's amazing and, this stuff's really easy to do. the knock on GraphQL for a long time is that it's incredibly powerful. But it's hard to get started with because you have to convert APIs into this GraphQL format before you can take advantage of all the stuff we've been talking about. That's not true anymore. You can bring any rest API to this world. You just write some configuration. I think for any developer, just trying this stuff out on your own API is, is worth the half hour. And it, it'll give you a, it'll give you something interesting to think about. I think that's a fair promise. And then, um, you know, look, I'll just go back to what we started with. I think been doing this a long time. I've just never seen so much energy. I've never seen [00:37:00] it across the developer landscape. I've never seen it across the, the, you know, C-suite in terms of the urgency of doing AI stuff. Like what a time. I don't know where this is going. I mean, I said some stuff. I think most of it's probably directionally right, but. The only way to really know and, and to find out is to get your hands dirty and shape the world. That's, that's why I do this stuff, and I just think it's gonna be a really interesting year to come.
[00:37:26] Andrew Zigler: I couldn't agree more. The joy of experimenting right now and being in the space, it's, it's a really unique time to be exploring what we can all build. And, uh, I, I know after this conversation I'm gonna go check out what y'all are doing. Um, there's been some really interesting tidbits about how teams can take advantage of this now.
[00:37:42] Andrew Zigler: And so we're gonna make sure we include this stuff in our show notes so that, you know, our listeners can go and check it out as well. And to you, our listener, if you've made it this far to the end of the conversation, then you clearly loved what we talked about today. be sure to give it a subscribe if you haven't already on wherever you're listening to the podcast.
[00:37:57] Andrew Zigler: Maybe even give us a, a like or [00:38:00] rating But if you are not reading our substack, which I mention every week, definitely go and check that out as well because we're gonna be dropping Matt's interview there. Along with the links and some news related to this topic, uh, because we cover things every Tuesday on our substack, so be sure to check us out.
[00:38:14] Andrew Zigler: And thanks for joining us today on today's Dev Interrupted. We'll see you next time.
