Matt
00:00:00.000
what AI really means is we're exposing these APIs much more directly to the end user. the APIs become the. Actual front door of the company, the interface of the company. It's not the website anymore. It's these individual capabilities that a AI model is going to make use of. And they were just never built for that. They were built for trusted engineers to use as a, piece of the internal implementation. This is very, very different.
Darin
00:01:28.491
Back in episode 3 21, we talked about CPS and earlier episodes, one of them being 3 0 6. We talked about GraphQL. I never thought we would see a mashup of these two things. Viktor, do you even see where this could potentially go?
Viktor
00:01:43.361
that depends whether we've wrapped m the agents inside CPS or not. But I'm curious. Let's see.
Darin
00:01:50.836
So on today's show we have Matt Delis on from Apollo Graph ql. He's the CTO. Matt, how you doing today?
Darin
00:02:05.506
okay, well, so we actually pulling the curtain back a little bit. All of us started, going back and forth of when you would come on prior to that time. So forgive me. That's, that's the problem. So back in July of 2025, he became the CEO, and again, I say congratulations, question Mark. Is that good?
Matt
00:02:22.931
No, it's great. There's, it's, it's an amazing time to be in tech and we're having a blast.
Matt
00:02:33.945
this is, the most meaningful change in my career and if you're excited about building stuff, there couldn't be a better time to do it. That's my view.
Darin
00:02:43.209
Hands down. I completely agree with that part, but after getting laid off in September wasn't my best time. So hey, but let's not talk about that. Let's talk about this mashup of MCP and GraphQL. I mean, if you think about it, most everything we've seen from MCP Viktor, correct me if I'm wrong, it's just been Jason back and forth, right? That's at the end of the day. It's Jason. we all agree about that?
Viktor
00:03:06.690
Most implementations and that I've seen, and I can go into a long run, why that is bad is just, translator to some API somewhere, right? Which makes you wonder why do we have a such an MCP, but basically that, that's what most CPS do, right? Just translate something coming from agent to. Whichever format API on the other end, uh, requires.
Darin
00:03:32.353
what's interesting to me about, and again, the, the pitch that came from Apollo was, you know, mixing the two together. It wasn't phrased that way, but that's the way I took it in. It's like, oh, okay. Well that sort of makes sense because graph. Ql. To me, the, the smartest use of GraphQL that I see on a day-to-day basis is GitHub, API. Right? That's a very basic, something I interact with every day and it's like, oh, it's just GraphQL. But now Matt, you're saying that GraphQL plus MCP is actually better.
Matt
00:04:02.473
Let me give you the context for GraphQL first. There's a technology piece to this, but if you zoom out, GraphQL is all about how do we build software that meets users the way they want to be met and the old days where you had one app and there was a pretty simple API behind that app that talked to your big database that's given way to a world where whatever brand you, think about, whether it's a media company or a retailer. Or, you know, travel, there's lots of different digital touchpoints with customers and each of them has to be specific to the, piece of hardware the customer's using or the use case. if you think about a hotel, you've gotta be able to make a reservation on the web. You've gotta be able to unlock your room with your phone. You've gotta be able to check in with the kiosk in the lobby. there's been this push to make great software that meets people where they want to be met. It and Graph Kill gives you a, a really flexible way to do that. it's a query language for APIs. So instead of having to build separate APIs for every different use case, you've got this common ground that you can start from and it lets, companies ship great stuff really quickly. And the reason we got into MCP, besides the obvious part where anybody doing anything now needs an agent strategy is. I think about MCP and I think about agents as just the next and most exciting chapter in what's a modern software experience gonna feel like, and, and how does a company reach its customers? Or how does a, a company reach its users? And I think million more that's gonna happen with AI and through agents. So we see MCP as a chance to go beyond the web and the mobile platforms that have driven API development for the last 20 years and, and have a whole new horizon of. Amazing things that are gonna get built, both, internally for companies and, now we're starting to see customer experiences that are built on AI and MCP. So for us, it's all part of the puzzle of how we help companies make amazing stuff and meet customers and users the way they want to be met.
Viktor
00:06:04.474
I can imagine that there are two very different use cases, and that would be for public and private APIs. Right. and I can easily see how something like MCP can help fit private APIs because. Uh, LLM is completely unaware of the existence of that something, But then when we talk about public APIs that LM already knows inside out. why would an LLM need an MCP?
Matt
00:06:33.487
one. I think way to look at MCP, this goes back to what you said earlier, is we just need a common standard, a common agreement for how models are gonna talk to stuff. And MCP wasn't the first take on that. Right. We've, had plugins and, functions and there've been different versions of this from OpenAI early on and then Anthropic shows up with MCP and I, I think the whole thing fit on a couple pages of markdown. It's a really simple spec and I think a lot of us look at it and recognize how simple and limited it is in a lot of ways, and I'm sure it's gonna evolve. we're in a lot of conversations about What the future of that looks like. But set the particulars aside. It's just a simple way to give the models, a door to the capabilities. Because if you want to build an agent, the agent is useful to the degree that it can talk to stuff and do things. An agent has to, whether, whether we're talking about, initiating a, a build in your CI system or. Managing a container, or if we're talking about putting something in your shopping cart and checking out, it's all actions that you want the model to be able to take. and so I think there's a lot of excitement early on around MCP and private APIs because companies see that there's so much opportunity to automate or, make intelligent these workflows and these systems that they've had for a long time. a agentic development I think has been at the forefront of a lot of this stuff because it's a safe place to explore and it's been a focus for anthropic and others to really, leverage models in that direction. But the same ideas I think are gonna show up in, in a lot of. Other places, like most people have an external API today, but it's not suitable to connect directly to a model. It doesn't have the right permission model, it doesn't have the right context limits. There's all kinds of things that you wanna change about that. And so, the thing might speak MCP, but most of the work we do goes into the question of like, what's the adaptation of one to the other and how do we bring more and more of the capability the models have to the world that, that companies have in terms of their, their systems.
Viktor
00:08:47.675
When you say like, right permission model, do you mean because the nature of permissions changes with agents or because, hey, I did not implement the right permission model, so this is my opportunity to not touch it and then do it elsewhere.
Matt
00:09:05.139
Yeah, I mean if you're a retailer, you have an API. Here's here, I'll give you an example. If you're a retailer, you've got an API somewhere that puts something in your shopping cart. But generally that's not exposed. That's not something a customer ever calls, and it's not set up to be externally visible that way, that API that's baked into your software, your mobile app or your website, or the other specific digital touchpoints that you created. When we talk about an agent, there's gonna be a different. Use pattern around that because by definition, the agent's not as deterministic. You might have different, models that come and go. Some of those might be the retailer's models, some of those might be the customer's models. what AI really means is we're exposing these APIs much more directly to the end user. the APIs become the. Actual front door of the company, the interface of the company. It's not the website anymore. It's these individual capabilities that a AI model is going to make use of. And they were just never built for that. They were built for trusted engineers to use as a, piece of the internal implementation. This is very, very different.
Viktor
00:10:17.248
Okay, that makes sense. If, if I understand what you're saying, basically we'll be talking about, hey, you don't go directly to it through the API, but you go through some kind of transformation within MCP. So the input, the MCP is very different from the output from that MCP that goes to that API. Basically, we are not, basically, we are talking EBS in MCP in a way, right.
Matt
00:10:41.903
hundred, percent. And, and, and sometimes I call this AI orchestration, right? We're gonna borrow a page from the world of, DevOps and, the cloud with other things. We've started to orchestrate, transformation's. One example, a great rest, API, you know, you mentioned GitHub. So let's talk about the GitHub, API for a second. GitHub's rest API returns an enormous amount of information. If you ask for a repository object, you get back. Hundreds and hundreds of fields, that makes sense for rest API. 'cause rest APIs are typically called by other systems. The latencies aren't that important. You can keep the parts you need and throw away the parts. You don't. If you think about a model, all that response is context. And you don't want all that context in the model that costs you money. It harms performance, it causes the model to hallucinate. one simple example. Of the difference between a great MCP tool and a traditional rest, API that was built according to the principles of, you know, resources and nouns and verbs and so on, is you're, you're gonna want to trim back to exactly what you want to feed the model, and you're gonna want to exclude the stuff that you don't. Or another example is, again, with GitHub, right? You've got a classic example of a, a rest API that's designed around that resource model. So there's a. Set of crud ops for repositories and another set of crud endpoints for issues and another set for users and pull requests and so on. So the actual interaction you might have with GitHub, if you're writing software against their API is like three or four or 10 different API calls back and forth. You don't want the model doing that. You don't want the model to orchestrate a set of calls. and so another example that comes up in this space is, is you need a. A layer that does that orchestration for you. Something that takes what you want to be a single semantic thing that an MCP tool exposes and, understands how to implement that as a set of sequenced or parallelized. API calls that hit the underlying resource.
Darin
00:12:41.702
You are using orchestration here? Uh, Kubernetes came to life because containers without Kubernetes was a pain at best.
Darin
00:12:50.792
Are you saying now that MCP tools without GraphQL orchestration. Is in that same level of chaos.
Matt
00:12:58.007
Yeah, and I think graph kill's a really early technology. So there's, there's gonna be a lot of change to this, but the idea, here's how I think about it. we're all about to be the proud owner of a lot more microservices and a lot more APIs thanks to ai because. If you ask an agent to write code for you, the best way to do that is to have that code be structured as independent modules and all the classic software design ideas. So Kubernetes showed up because you don't have two containers. You have 2000 containers. And there's a scale to the problem when The need for some kind of piece of infrastructure where you describe intent and the infrastructure figures out how to make the systems configuration match your intent. That model made a lot of sense. I think we're gonna see the same thing for APIs. the idea today that APIs are, you know, in, in DevOps we have, Cattle not pets, right? we used to have, individual Unix machines and you had your, Sendmail server on one Unix machine and you've got your, database on another one. But now we just put them all in containers and the containers come and go. I think the same's going to happen with APIs like today. They're handwritten, they've got names. We think very, very carefully about their lifecycle but I think tomorrow you're going to see a world where a lot more of that's ephemeral and you're going to want a piece of infrastructure that orchestrates them the same way we do with containers. and GraphQL is, is an example of a technology that, that's built that way. It's a declarative technology just like Kubernetes, where you, you describe what you want and the system figures out how to do the actual imperative steps to accomplish that.
Viktor
00:14:31.416
Who would be writing that from GraphQL? Because earlier you mentioned, okay, so you don't want AI to make five calls probably because of the waste context
Viktor
00:14:43.596
Yeah. But then kind of can you make it dynamic? Okay. You, you made these five calls first time kind of like save it. Save it as GraphQL or, or whichever other script or whatever it is, kind of like, okay, you did it once. That was inefficient. Kind of like to catch it.
Matt
00:15:01.621
Well, to me it's not an efficiency thing as much as determinism. Let's say I'm a bank and I want to have an agentic way to look at my bank account. I wanna make sure every customer sees the same thing, right? It should be five recent transactions, not 10, and should have these results that come back. And You don't want the model kind of freelancing how it's gonna combine different systems together to give a user an answer. actually want a consistent experience and you want to be able to stand behind the. Performance and, you know, reliability and, and all the other stuff that's so important.
Viktor
00:15:33.371
here's one thing I don't get really. don't we already have reliable, deterministic waste? That's basically everything before AI isn't actually non-determinism. The, the strength in a way, right? Because I can just say, Hey, show me my buying bank details and order them randomly, because I, I'm just silly today. And, uh. Divide them all by 75, I dunno. Right? Kind of like completely non-deterministic, right?
Matt
00:16:04.878
Yeah, but I think non-deterministic models pair well with deterministic capabilities. They can access, you don't want non-determinism all the way up and down the story.
Matt
00:16:26.400
Yeah. But take Kubernetes as an example. If, the action I want to take is to schedule a container, I want the rest of that story to be precise and deterministic
Viktor
00:16:36.657
Yeah. And that's the part you don't. To involve AI in at all. That's what I'm trying to distinguish, kind of like why would we involve, so there is a world in which AI uh, lives together with deterministic parts, and sometimes we use only deterministic things. Sometimes we use only AI or people. Same thing, Sometimes we combine them, but kind of there is no reason to, if you would, for that specific case in Kubernetes, if you would bring ai, probably the answer would be why.
Matt
00:17:09.779
Yeah, and that's, my point is you want both. you want the non-determinism, the freeform experience that AI makes possible and you want underneath that it to rest on a bunch of precision that I can observe, I can secure, I can, manage in various ways. That's the deterministic half. And my argument is just in a world where you've got ai, where you're gonna want to. Iterate really quickly on those experiences where there's probably an exploding number of systems underneath the hood that your ag agent coding, investment has produced. It isn't gonna be enough to hand write the, kind of sequential code that you no used to write, just like it doesn't make sense to hand write the container scheduling you used to write back in the day. You're gonna want a system that does that for you and has a, a behavior you can understand and manage.
Darin
00:18:01.582
You are talking about these exploding systems. I mean, to me that rings platform engineer all day long. Uh, what's my first step? Am I gonna rip out all my existing infrastructure now that I've got all this new stuff that's like I gotta put in?
Matt
00:18:14.707
I think AI platform engineering is gonna be a thing, and I think it's gonna be that, that combination of, gonna be where the water's edges between the non-determinism of the models and the deterministic systems that you can, bring to bear, but. AI touches so much stuff, right? I might have a deterministic execution story, but I'm gonna want to use AI to digest all of the traces that come out of my observability system and make, recommendations, about where performance might be improved. there's gonna be a whole new practice, I think, of people that come with a systems and platform mindset. But apply the modern, AI capabilities to each of those pieces. And so how much that means the Stack is gonna change and how much it just means each of the components is gonna evolve. I think we'll see, case by case. but I think the, the overall all, footprint for a modern platform org is gonna look really different in a couple years from what we have today.
Darin
00:19:20.605
Well, let's flip it a little bit. Let's say I'm not the platform engineer. I'm just somebody that's using, you know, I'm a developer using cps that's just wired into my log code or whatever tooling I'm using. this is a mic. For me in some way, shape, or form, right? Because now things that used to not be deterministic, because the MCP was sort of chatty, for lack of a better term and never ended up at the right place, now can become either fully deterministic or at least more deterministic. I mean, does my workflow as a developer change.
Matt
00:19:51.570
Well, it already has, right? I think like we've gone from everybody doing the work by hand to a world where you've got, Claude Code or, cursor or whatever your, favorite tool is helping you save time by pre-writing a lot of the lines of code for you, to a world where now it, it knows how to go kick off a build or, or push a, gi change to a world where we've got autonomous agents, where the developer is really kind of kicking off a, a series of tasks and plays a, an architect role, but the end to end is being done more and more autonomously, by the systems. And the thing that makes all that work is, number one, you've got all your systems set up in a way where the AI can talk to them. in the MCP world, it, it doesn't work that well to have a hundred MCP tools. You know, we're, we're talking to a lot of teams that are solving the problem of, well, how do I, how do I take all the different stuff that I've got in my organization and bring it all together to developer? And you don't want to ask each developer to solve that on their own. So there's your classic, platform engineering task and, and a mindset of, serving others in the org and making them more effective and, able to focus on the things that they're tasked with and not the undifferentiated stuff underneath is really valuable.
Darin
00:21:04.170
Use the word, we're becoming more architects than we have in the past. I don't agree with that. I think we're becoming more product managers than architects.
Matt
00:21:25.333
Yeah. Well, no, no, no. I think there's two specific things that, um, come together really magically. If you've got modern AI tooling, one of them is you've gotta know what you want, so you have to have some kind of a product. Vision. that's the product manager piece. what we're finding is to use them at scale. it's incredibly valuable to have a systems mindset. 'cause I, I can, I mean, my experience is if I understand how the different components of a piece of software I want to build. Relate to each other. If I have a point of view on the interfaces or the APIs or the database schema or whatever, whatever your interface layer of choice is, I can tell the model what I want and I don't really need to pay too close attention to the particular code that the model writes because I've got that, system orientation about I know where the boundaries and the interfaces are, and if I can reason about those, I'm in good shape and I can trust the model to get The detail, right. For each, each unit or each component of that. And I feel like people that got both of those, that have the, the systems mindset and that have the taste or the creativity, the product piece, that's where the magic happens because those are the two things that, guide the model toward a, an outcome that you're gonna like.
Darin
00:22:51.386
Okay. Sure. Okay, so that's the first part. So I wanna say, yeah, because if you said no, then I wasn't gonna ask the second part. The second part is it feels like maybe then we're trading one set of problems for another because magic hand wave, I've got my quad code set up and I've done the exact wrong thing. As you said earlier, I've put in 175 CPS into my cloud code, and now I'm needing to mix in. GraphQL on that. I mean, I'm not doing it, but as the MCP author, I'm now mixing in GraphQL potentially. Is that the implementation Stack that you're looking at? How, I mean, how do you see it?
Matt
00:23:29.631
Well, the big advantage of GraphQL. For an AI model is that it distills a lot of complexity into a query. It looks like a SQL query, right? It's, it's 10 or 20 lines of, GraphQL. The complexity moves into the guts of your GraphQL implementation. So just like you have a, a database engine that's got an incredibly complicated query planner, you've got a GraphQL piece of infrastructure that knows how to turn that query into all of the. API calls and transformations and, joins and so on that you might want to do so why is that good for a model? Well, now the model's writing something really concise. Instead of writing a thousand lines of procedural code to do what you would otherwise have to do, it's the same argument for why a model can write a SQL query. It's the same argument for why a, a next JS component makes more sense than having the model write bear. JavaScript that runs in the browser, like these abstractions save a lot of complexity from the, context budget, and they move that complexity into reusable infrastructure that typically you don't have to think as much about. So I just think of it as moving the, bits around so that we can take best advantage of the model if you're doing ag agentic development. And your entire surface area of APIs in your company is visible through this GraphQL schema. The model is so much more effective than if anytime you want to do something, one of the steps is you're, you're building a new microservice that joins all that together and now you've got a whole, DevOps life cycle around like deploying and, and testing and. A runbook around that, that's the complexity that we don't think about in the software cycle that actually adds up. I think over time, that you can replace with infrastructure.
Darin
00:25:24.862
You were saying concise and context windows, and we're talking about GraphQL typically coming back from GraphQL is Jason format. Right. That's, that's what we see
Darin
00:25:38.062
all the time, probably. But then there's been this new concept called Tune. I'm assuming you've heard of tune, the token oriented object notation. Right. Does that even come into play now? I mean, is that the next big wave? It's like, okay, now, because in theory, Jason was created for us humans.
Darin
00:25:59.002
Uh, we don't need as much. Excuse me. The machines don't need as much context as we as humans like to have. I can argue tune is readable too, but let's, but it's much more compressed than what a Jason Blob might look like.
Matt
00:26:14.842
Yeah. I think there's a bunch of different takes on this, but there's no question that if the output is destined for a model, they're better in coatings than your traditional. You know, Jason or XML or whatever you might have had in the past, and I think we'll see more on that front. clearly though one of the big benefits of the model is how good it is at, reading unstructured information. And I, I think historically APIs were built, against a very different use case. So we should expect to see new things there.
Darin
00:26:43.756
Do you think that will come sooner than later, or do you think it'll be much later?
Matt
00:26:47.176
A lot of people are already doing it, right? that's an easy change in most stacks to make. And it's, it's an area of active experimentation that I know a lot of people are excited about.
Darin
00:26:55.646
You moved to CEO, now, eight, nine months ago. did they take away all your development credentials as you went to that level?
Matt
00:27:05.196
I actually do more. development now than I was before. That's because of ai. one nice thing about agentic development is I'm not limited because there's one or two technologies in the Stack I want to use that I haven't been able to come up to speed on. That was always a big stumbling block for me. I mean, I mentioned Next.js before. That's a good example. my JavaScript background is, Way before modern React and modern, Next.js But with models, that's okay. I don't need to spend my evening coming up to speed on all the new, hooks, interfaces in React or whatever. it is this year, the models have, gotten really good at writing that kind of code. that's, let me actually do a lot more. I, find it really empowering.
Darin
00:27:52.188
The interesting part to me, what you just said is now you have a C level. That was a CTO and now was a CEO that is enjoying coding as a CEO. Would you recommend that to other CEOs? if they're saying, okay, AI's gonna solve everything. So now we're doing AI all throughout the company. So everybody's starting to do it, but they haven't tried it out themselves to see, oh, actually it may actually make a real difference. Not just because they went and play golf with their best friend and now we're gonna AI the things I.
Matt
00:28:30.049
there's so many angles on this, but they all point to the same answer, which is, yes, yes, yes. You should write, code, and you should use AI to do it. one angle is if you're not excited about what this all means, then I, I don't know what to tell you. Like I think AI is going to change everything about everything, and. one of the places for structural reasons that's coming first where AI is already incredibly transformative is software development. and that's because, and Anthropic and, and Google and, and others have made it a priority and they've, they've focused their efforts in that direction. So I think it's the best way by far. To really understand just how powerful this stuff is, much more so than, cracking open GPT and just like talking to it or asking it to, to write some text for you, and that can be eyeopening and inspiring. I think another angle is, for a company like Apollo, if I'm not conversant in, today's world of how development gets done and, and what the pain points around that are, then I'm not doing my job. So. it's about being immersed in our own world and understanding it. And then I think the third thing is like, I think this is true for any job, but certainly for mine. Like there's so much that a two, three years ago would've taken up a chunk of my time that can be automated or improved with ai. I've gotten. Incredible benefits from, this kind of stuff ev everything from, doing research to digesting a lot of what, everybody in the company is doing and, and helping me have a, a better view on it. I think some of that shelfware where you can just, download the product and use it and you get the benefit. But it turns out there's a last mile of integration, of, of building simple agents that combine things together they're not outta reach, but you've gotta want to do it. And if you do do it, you have an enormous advantage. And it's, it's one that I think is essential now. So I see it, take your pick as, a skill that, everybody should develop and have, and certainly for the work we're doing, I, don't see any other option.
Darin
00:30:35.373
I think the other side for especially the C levels, is maybe not so much code they can if they want, but modeling what changes you wanna do in the business. Develop a handful of different agents that model, you know, maybe you're wanting to get into a new part and it's like we don't know much about it. Great. Go create an agent to start doing all the, doing the work and then having an aggregator to pull it together and that way you get your exact summary and off you go.
Darin
00:31:01.422
let's rewind back to the, the implementers, the, the people, as I like to say, that are stuck doing the work. What we talked a little bit about what should they be doing now? here we are, it's in March of 2026. What should people be doing right now?
Matt
00:31:21.033
My angle on, really the whole company starts with the software should serve the user. So I think there's a lot of exciting opportunity if, you're a, business that has customers that touch you digitally to think fresh about how that can and should feel. I think there's gonna be a lot of room for innovation the wonderful thing about AI is it's really easy to experiment. So there's a lot of different things that can get put in motion. I think the companies that are thinking most thoughtfully about this have a, culture they've deliberately tried to create, not just around experimentation, but a, culture that expects a lot of work to, compresses the schedule and, and expects a lot of the work they do to be, replaced or subsumed in, in just a few months by a new generation of, of AI technology. teams that were building big rag systems a year ago, for example, shifted over to building on MCP. And I think teams building on MCP today are gonna, find another shift in their future too. The old stuff doesn't go away, but if you come in with the mindset that you can have a waterfall plan or you can, build for a three or five year horizon, I think you're gonna find some uncomfortable truths as you go forward. So I'd start with just making sure you're in the right mindset and you've got the right culture in place. and I, I think it's a, a time when moving fast and, really an ownership and empowerment mindset gets you a long way.
Darin
00:32:50.508
So moving fast in a waterfall method. The way I see it because it's gotta, if there were, it's ever a time for Waterfall, now is the time.
Matt
00:33:00.378
I think about it like more like platforms. I think. Here's the puzzle. Everybody is urgently trying to navigate how to build for an AI world. That's a must do, but. You know that you're not on stable ground, you know that a lot of things are gonna change. think about just on MCP alone, the pace of change of that spec and how different it is from when it first came out. to me, it points to platform engineering because you're only shot as a company if this outcome really matters, is to get the foundations in place. if you don't do that, you're gonna end up with something that, half works. You're gonna end up with an enormous security hole or some kind of performance issue. You can imagine all the problems that come up and, and I, I just think that's, it's doomed from the start If your, strategy is just, eh, let's have all the engineers play around with a bunch of stuff. so that part of it, I think you have to have a. a view, but the view has to be anchored less in, in like a, a very specific technical, direction and more in a, a mindset of what are the building blocks that I'm gonna need and how can I arrange those building blocks in a way where I've got a path forward that lets me, you know, explore quickly. And when I find something I like, I have a, I have a path to get that into production and I'm gonna like where I land when I do that.
Matt
00:34:25.614
I mean, I'm, I'm always a, platform guy at heart, so I wouldn't fight that, but I think there are chapters when it's a lot less important in some areas than it is in others, and, and I just think we're in a moment now where it's essential.
Darin
00:34:39.068
Who are the people that shouldn't be even bothering with this right now? Are there any.
Darin
00:34:46.553
Yeah. This meaning doing the AI thing or MCP, like ignoring governments or anything else, it's like, are there any valid business cases today where AI should not be coming into play?
Matt
00:34:59.708
I have a hard time imagining as a user, I have a hard time imagining that I'm not gonna want and expect. an AI first kind of experience with almost every company that I interact with. I just think about how quickly the models have replaced some of the things I, I used to do a different way, it seems, unwise for I think most companies, to not have a strategy for how they're gonna. Offer that kind of experience to people. I'm sure we can come up with an exception, but I think they're gonna be few and far between.
Viktor
00:35:39.364
I am sure we can't actually, I mean to me that feels very similar to conversations we had a long time ago. You know, when we started the internet and all the jazz and then many companies were, oh no, we are a bank. Kind of like, you know. We don't do things online, kind of like who's going to do it online? Come on. Right.
Viktor
00:36:01.894
Those conversations existed. Now, probably everybody would deny that they existed, but they, they were there, right? Most companies were, not most, but many companies or groups or companies for No, no. This is amazing. Kind of like my daughter uses it, but come on. Not for business.
Matt
00:36:20.239
Banks are such an interesting example because, here's what I hear. the challenge if you're a retail bank is that you've got a lot of customers that have very different ideas of what you really are, right? You, you've got retirees who think about their bank account as, you know, kind of a classic thing, and maybe they're balancing their checkbook every month. You've got. You know, millennials who think about their bank as, a set of like financial partnerships and they're trying to mass credit card rewards points. And if you crack open a modern bank app, you'll find a whole bunch of call outs to, a whole view of a financial portfolio of stuff. And if you're, you know. In your twenties, you have a, completely different idea again of what a bank even is. And somehow a, a bank has to meet all of them in one place. And that's really hard to do with a rigid app in a, old world pre ai. But what AI means is now, each person's experience can become personalized and it can adapt to what that person most wants. That's really exciting, right? Like, that's what I want. When I interact with an org. And so I think it's inevitable in that space, that AI's gonna have an enormous impact. still a bank under the hood, right? Like, I mean, hopefully they're not gonna lose your money. But the touch point is, so d.
Viktor
00:37:37.125
In case of banks, I feel that, feel that we are gonna go back to where very old people are today. younger people, I mean, even middle-aged people. You know, you go online and you check your stuff. And then if you're 80, you go to your, to your bank, you sit in front of the clerk, you explain what bothers you, and you get a humanlike answer, right? Rather going back to that anyway, except that clerk is not anymore, real person.
Darin
00:38:08.787
Everything you're saying there resembles me, so be very careful what you say next.
Viktor
00:38:13.300
are you going, are you going whenever you need to do transaction, do you go to your bank offices? Darin,
Darin
00:38:23.305
old enough. I think that's the key word. I'm old, not, not enough. I. One thing I heard back in January we're talking about the ai, we got a habit type deal. there was a company Tailwind, the company behind Tailwind CSS,
Darin
00:38:39.310
back in January, had to lay off 75% of their people. Their revenues are down 80% because right now you go to AI and he can tell you how to do tailwind.
Darin
00:38:49.900
Are you concerned about that as a business? I'm comparing apples and not even oranges, you know, I'm comparing it to pine cones or something, but I see tech companies to where I'll use me. My language of choice is Java still, it's just, it's been 30 years. That's my language. But in all the AI stuff I'm doing, it's 50 50 go or TypeScript. And I don't know how to write either one of 'em. I can read both of 'em, but I don't know how to write 'em.
Darin
00:39:22.251
exactly, but this goes back to the architect versus product manager versus all those things. It's like, okay, I know what I need to do. I know I need to do security audits. I need to do all these things. Right? So having been around 40 years in this game, I understand those pieces. should tech companies be really concerned? About this, especially if it's like, oh, we need, we need to send 'em to our docs page for everything. Do you.
Matt
00:39:45.921
Yeah, exactly. Like if your business model relies on human developers going to your docs page to figure something out, and somehow that's the top of your acquisition funnel, I, I think you've got a problem. and I think that's true beyond development companies. think about what SEO meant for so many different kinds of companies. And now we've got an even more, salient issue where if a customer is gonna find you or not, depending on what the model says, when they're typing, it's going to dramatically change the landscape in so many areas that seems clear. I think it's gonna raise interesting questions about what their advantages are, what the competitive advantages, in that landscape. Like if I'm a retailer, is my advantage a about better product catalog? Is it a better user experience while buying? Is it a better story around quality or support? You know, there's a lot of different angles on it, but you're gonna have to reckon with. A major shift, I would think, in how people find you and, how you convert those into purchases. Same's true for developer tools. No question.
Darin
00:40:56.122
So a polygraph ql, that's so much easier to say. Why is it so long? Why can't it be shorter? Matt can't, can.
Darin
00:41:04.567
apollo.dev. Thank you for me. So the fic, so Apollo do dev is what's, what's that site about? Since we're just talking about sites that may no longer need to be there to begin with, can I find everything in my LLM that's on Apollo? Do dev right now.
Matt
00:41:18.367
Yeah. Yeah. So, so like, like any, so of core piece of the modern Stack, one of the things that's really important about Graph Hill is the models are good at it. And certainly when we write. Technical documentation. Now, our, first audience in mind are the models that are gonna get trained on that. and I think that's like a shift in mindset. We used to talk about developer experience, now we talk about agent experience. So when we build a a product, we have to make sure that agents can interact with that product. humans are still gonna use it too. We have to be mindful of a world where we're, we've got one leg in the. Present and one in the future. But we have to think first and foremost about the AI first take on all this stuff. And, and you see that across the whole landscape with, development companies. And, I don't see that not continuing.
Matt
00:42:16.264
Well, MCP kind of was the trigger because that started, I think I, I guess it's a lot of things together. I'd say early 2025 is when a few things came together. One is the foundation models themselves got good enough to write good code, and the other is. Again, these agents are only interesting to the degree they can talk to stuff. And that's when we started really seeing the first end-to-end stories where, it's not just pre-filling a, function definition or a unit test harness for me, but it's, it's actually, running unit tests and interpreting the error results and then making the code changes that correspond to that. It's actually submitting a pull request. a lot of the companies like ours that have ACL I, as soon as the agents had a way to run bash commands that unlocked a whole bunch of workflow that you could do. and then I think what's happened in the. year and a half sense is you just get a lot more recognition of details to make that really delightful and get it right. Right. So a lot of the, iterative work on feedback loops, I'll, I'll give you an example from GraphQL. GraphQL historically has had a pretty weak story on errors. I won't bore you with the details, but. A GraphQL query by its nature can be pretty flexible, but that means there's a lot of different things that you might have to report back, that didn't work. Maybe the upstream API wasn't available, for example. we're working on some changes to GraphQL that, make those errors more clear to what we now assume is an agent that's interpreting the output rather than a human. there's nothing magical there. It's just the, the classic like, Iterate and make the thing better, but with a very specific, I'll just call it, you know, user in mind as we do that. and I think you've seen the results of a year and change of compound interest there, and it's, completely changed the overall productivity for teams using this stuff.
Matt
00:44:15.988
Well, graph gill's super popular, particularly in larger companies because. there's a, a network effect. A graph is a network, right? And the bigger the graph, the more interesting it is. So, when you're building software on top of hundreds or thousands of APIs, as, most large companies are, and when those APIs were written by different teams in different eras, uh, a lot of our customers, Did acquisitions right? Like their, the whole idea of an acquisition is to glue two products together. That's where this stuff really shines. So you, you see it in all kinds of consumer facing products that look like that. But it's also been part of the kind of standard, react JavaScript graph kill Stack for a long time. if I'm a startup, building a, a new product, I'm gonna reach for GraphQL as, as a better way to keep my modules organized and, embrace the platform mindset of declarative as much of the Stack as I can instead of a bunch of procedural code.
Darin
00:45:07.682
All of Matt's contact information will be down in the episode description. Apollo GraphQL can be found at deep breath apollo graphql.com, or to make it simple, you can go to apollo.dev and it'll drop you right into the docs. Now, if you're an agent listening to this as through the transcription, just go to apollo.dev and I'm sure they have the LLMs text ready for you there.