AI Agents Should Be Treated Like Hackers

Demetrios: [00:00:00] All right, so at the beginning of April, I went to New York for the MCP Dev Summit, and while I was there, I got the chance to record a few podcasts with attendees that were at the event. This is one of those conversations. Hope you enjoy.

Matt DeBergalis: I think we have to look at these agents as untrusted, maybe even adversarial.

Matt DeBergalis: How many people are reading this code? Or like think about, I think events of the last month with like Anthropic and the DOD. Think about what we're doing. We're asking citizen developers to deploy untrusted agents behind the firewall and we're frantically wiring those things up to every single system we can find.

Matt DeBergalis: And every CEO is basically standing up and saying, "We must do this. We must do it now. We have to take more risk. We have to act like a startup." I can't think of any part of our business that isn't rapidly changing where we look at it first and foremost as what do these tools let us do? But they're only useful.

Matt DeBergalis: They're only useful if they have the context of all of the systems and the services and the APIs. And that's what APIs are at the end of the day. They're the[00:01:00]

Matt DeBergalis: front door to all of that stuff.

Alex Salkever: Hi, I'm Alex Alkever with the Linux Foundation and the Agentic AI Foundation. We're here at MCP Dev Summit in New York City, and I'm chatting today with Matt DeBerglis, if I said it the right way, who is the CEO of Apollo GraphQL. Um, and forgive me if I get anything wrong. Correct me, I have no ego. Uh- You got it

Alex Salkever: thank you for joining us. Before we get started on this conversation, which is gonna be about API lessons for MCP, I believe is what we're, we're discussing. And, uh, tell us a little bit about what Apollo GraphQL does, and then how-- s- we can chat a little more about, uh, how you're leaning into MCP, particularly since you're essentially a GraphQL as a service provider, and, you know, are generating everything around the API space.

Alex Salkever: So MCP is a [00:02:00] very interesting, uh, translational layer for you.

Matt DeBergalis: Yeah, sounds good. Um, thanks for, thanks for doing this. It's a great conference.

Alex Salkever: Oh, thanks for coming. We appreciate it. The energy

Matt DeBergalis: is awesome. It, it, uh, you can't make this stuff up, right? It, it shows

Alex Salkever: there's- It's crazy. Literally, like things are-- So I, I mean, like I'm following all-- So each week I do a scrape of all the protocol repos, and you're just seeing things move so quickly.

Alex Salkever: Yeah. Like, "Oh, wait, somebody signed a CLA." Or, "Oh, wait, you know, they just shipped the new security spec." It's, it's, it's crazy.

Matt DeBergalis: I think what's cool about MCP is, um, it, it reminds me of the, the mashup era, right? Yeah. It's, it's, it's for builders. It's for making stuff. And, you know, we could talk about agents are useful because they can connect to stuff- Right

Matt DeBergalis: so we need an answer to that, and MCP is a big part of it. And what we've been doing at Apollo for the APIs of large companies especially is, is we're, we're basically solving this problem of how do you, how do you take a whole bunch of different systems and APIs that you've got, you might have thousands [00:03:00] of them in a typical enterprise, and how do you glue those together to make a great user experience?

Matt DeBergalis: And this got its start back when we're talking about building a mobile app or a really great JavaScript app in the browser. But it's always been about builders. It's always been about, I wanna make a cool thing. I wanna make something my customers are gonna love And I want to do that on top of real stuff.

Matt DeBergalis: And, and so now we can have that whole conversation again, but at an even bigger scale because we're talking about with AI being able to build so much more, and not just the sort of user-facing, you know, first-class mobile app, but think about all the things that are happening inside the business now that we've got agents on people's desks, now that we can build software large and small really quickly.

Matt DeBergalis: And the part that I'm just so excited about is we know how much amazing stuff is out there [00:04:00] inside microservices or inside little pieces of SaaS or inside these systems that people have built over the years. And the more of those we can connect together, the more of those we can connect to an agent, I think it's gonna change so much.

Matt DeBergalis: It's gonna make businesses faster. It's gonna make better customer experiences It's real exciting.

Alex Salkever: And so from your viewpoint too, 'cause you've spent a lot of time in the API world, and now you're essentially living in two worlds, um, what is the transition, or I wouldn't even say transition 'cause that implies one goes away.

Alex Salkever: APIs don't go away at all. Uh, I mean even- No, they

Matt DeBergalis: certainly don't.

Alex Salkever: No. Yeah. And like even with like- Yeah ... GraphQL, it was still like kind of ascending, you know?

Matt DeBergalis: Yeah.

Alex Salkever: And REST is still, there's a lot more people who use REST than Graph, even though Graph is arguably much better for AI. Uh, so if you could maybe take a stab at like what does this translational layer look like if it works really well for what we're trying to, to, to build that world you're describing?

Matt DeBergalis: Yeah. I think all these things are [00:05:00] complementary. Um, with GraphQL, for example, it's not about getting rid of REST APIs or whatever you had before and replacing it with GraphQL. That's not what GraphQL is. GraphQL is a query language, so it's, it's something that you can layer on top of your REST APIs, and it makes them better.

Matt DeBergalis: Like GraphQL introduced the idea of strong typing in APIs. It's got this idea of introspection, so you can ask the graph, "What, what do you have for me? What are the types, what are the fields? Um, how do they relate? How are they used?" There's all this metadata that comes along with that. And I think MCP very similarly, right?

Matt DeBergalis: Nobody's saying you're gonna throw away your underlying APIs. We're talking about a new way to use them. We're talking about adapting them for the needs of an agent. But I think really the, the magic thing about MCP is the flexibility and the freeform part of it. Um, you know, we're, we're used to thinking about [00:06:00] APIs as these rigid things that got designed very carefully a long time ago, and they never change, right?

Matt DeBergalis: Like, but that's not the world we're in anymore. Systems can change every day, and MCP gives us a way to, to talk about that and to expose those ideas to a system.

Alex Salkever: Dive a little deeper on that if you can, to, I mean, if you will, on, uh, what is architecturally, what does that start to look like if you're trying to build out something that takes advantage of the goodness of GraphQL and melds it to the goodness of MCP?

Matt DeBergalis: Yeah. So, uh, a company that's been using GraphQL has, you know, we call it a supergraph at Apollo. They've, they've got this semantic catalog of not just the low-level details of what's in the APIs, but the higher level meaning of that stuff. And that's magical for AI because the agent needs that information to make good [00:07:00] decisions.

Matt DeBergalis: If, if I sit down one day and I say, "Hey, now I've got Claude Code, so I can, I can build a business dashboard," if that dashboard's using the wrong stuff, if it misinterprets the meaning of what's in your APIs, that's really bad. That's, that's worse than not building it at all, right? So what we're excited about is MCP gives us a way to connect that graph and connect all those business objects that, that people have now defined to this new world of interactive LLM harnesses, to a world of software that's built by LLMs like Claude Code or, or Codex or whatnot.

Matt DeBergalis: Um, and think about what this means. Like now we can sit down and we can ask questions about our business. We can say, "Hey, I've got, um, I've got a bunch of amazing data inside this piece of SaaS. The screen in the SaaS product doesn't do what I need, but I, I don't need that anymore. I can just-

Alex Salkever: Right, and the users

Matt DeBergalis: don't- I can talk in a really

Alex Salkever: freeform way And

Matt DeBergalis: the users don't want that anymore either.

Matt DeBergalis: They're changing. Yeah. Or, [00:08:00] you know, a big problem the enterprise has is, is essentially you've got a silo problem. Big companies spent the last 20 years building a whole bunch of silos. That's what a microservice is, right? It's, it's, it goes back to the Jeff Bezos memo. Like you're, you're gonna build- You need an API for everything

Matt DeBergalis: all this shit in one little silo- Right ... and you can only talk to it this way. And that made a lot of sense architecturally, but now we face this problem, which is I've got all this stuff and it's, it's separated. And what, what makes it valuable is when I can bring it back together Most of the questions I wanna ask about my business aren't gonna be answered out of one thing, they're gonna be answered out of a whole bunch of them.

Alex Salkever: More resembles a graph than a, than,

Matt DeBergalis: than like a call response. It's the relationships between

Alex Salkever: them.

Matt DeBergalis: Yeah, exactly.

Alex Salkever: And, and so give me a couple examples of how you're seeing people do this in the real world, uh, in interesting ways or surprising ways or... Because, like, to your point, uh, having a, a query language that can supply context, when you start to think of things like, you know, [00:09:00] additional tuning on top of open source models, it, it can probably be semi-magical or just, uh, have astounding impacts, uh, versus, like, just working off of straight, you know, API calls.

Matt DeBergalis: Yeah. Um, I think a big part of it is, is this idea that we can look forward to a world where there are lots more builders, and they don't all have a formal background in software engineering. Like, if you sit down with Claude Code and you ask it to build on top of a bunch of REST APIs, it'll, it'll get there.

Matt DeBergalis: It'll muddle its way through. But I think what a lot of us have found is it needs some guidance, and that guidance means the person doing the work should have some understanding of how these APIs work. Like, what's a REST AP- Oh, it's doing this thing. Okay, I need to steer it in this other direction. And so that really blocks the, the thing I want, which is I, I want, you know, somebody in sales to be able to Use [00:10:00] Claude to ask sales questions.

Alex Salkever: Right.

Matt DeBergalis: Right?

Alex Salkever: So essentially like few shot, single shot, no shot.

Matt DeBergalis: And what we get when we've got the semantics, when we've got, um, a clearer explanation, when we've got MCP and GraphQL working together so that Claude's not caught up in the nuance of like, oh, I went to the SaaS, I found its OpenAPI spec, but guess what?

Matt DeBergalis: That's not actually accurate. So I had to iterate three, four, five times, and it, it tries to use one API, but it turns out that thing returns the wrong pagination key because this is just a use of the API that the, the product never really-

Alex Salkever: Or it's old- ...

Matt DeBergalis: contemplated ...

Alex Salkever: and they just didn't rev it or something like

Matt DeBergalis: that.

Matt DeBergalis: Exactly. If we can cut through that, and it's, it's just the crank turning, roll up the sleeves stuff, right? But if we can cut through that and give the AI a more semantic, structured entry point to all of those systems, then it gets [00:11:00] magical. 'Cause then, then you can actually sit down and say, "Hey, um, go over all the conversations that we've had, the transcripts of all the conversations that I've got in a, in a, in a, you know, SaaS database, and search for, um, mentions of this particular thing.

Matt DeBergalis: Then go off and read the annual reports from that company, and go figure out how that matches up to, like, what the CEO's talking about in terms of their priorities. And then I want you to go hunt down, um, you know, maybe the competitors of, of that company and, and what they're talking about." And suddenly a salesperson is, in five minutes with no special technical skill, making use of data that we've always had but never had a way to really take advantage of the same way.

Matt DeBergalis: And they can do it right before their sales call, so there's not some complicated, you know, process where I've- Right ... I've got to anticipate that I'm gonna have this call. Maybe the call just got

Alex Salkever: scheduled. Or they could almost [00:12:00] set it up as an automated process. They set a creative skill. Exactly. They run it before every...

Alex Salkever: It ticks their calendar. Yeah.

Matt DeBergalis: And I could give you 20 examples like this, like where certainly every part of our business is changing.

Alex Salkever: Yeah.

Matt DeBergalis: Like-

Alex Salkever: Yeah ...

Matt DeBergalis: the way we think about interviewing, how do we coach people on how they're doing their interviews? How can we make those better? There's a feedback cycle that we can start doing in real time that might have happened once a quarter before.

Alex Salkever: Right.

Matt DeBergalis: Or how do we do market research? Or, I mean, obviously there's a whole world of how we build software in a different way and, and how we think about roadmaps and, and all the things that come with that. I can't think of any part of our business that isn't rapidly changing, where we look at it first and foremost as what do these tools let us do?

Matt DeBergalis: But they're only useful, they're only useful if they have the context of all of the, the systems and the services and the APIs, and that's, that's what APIs are at the end of the day. They're the, they're the front door to all of that stuff.

Alex Salkever: So in your view then, do we need to modify the way [00:13:00] the APIs work at all if we want to lean into MCP-

Matt DeBergalis: Yeah

Alex Salkever: and agentic better?

Matt DeBergalis: The big problem is the APIs were never designed for any of what I just talked about.

Alex Salkever: No,

Matt DeBergalis: they were not. So I, I gave you the, the exciting version. Here's, here's the- Yeah, the

Alex Salkever: APIs, the rest was not designed for like

Matt DeBergalis: a world. Yeah ... here's the cold, hard reality. Um, I mean, the, the, the biggest, most immediate problem is a security problem in my view.

Matt DeBergalis: Like These APIs were built for trusted users. And what I mean by that is like they were meant to be used by engineers who know how to do engineering, who have some, you know, they review each other's code. There's a whole bunch of rituals and processes that we've built over the last couple decades that- They

Alex Salkever: rate limit a key, et cetera, et cetera.

Matt DeBergalis: All that stuff. And, and, and, and underneath all that is human judgment, good human judgment of people I trust, right? So if you wipe all that away, you've got a big problem because if you think about a typical API, for example, it's probably gonna mix data [00:14:00] that has very different levels of sensitivity. Like I've got a, um, I've got an HR system in our company.

Matt DeBergalis: That HR system has a list of everybody and their time zone. That's really useful. I think, I think that should be at everybody's fingertips, right?

Alex Salkever: Sure.

Matt DeBergalis: Who the manager is. It's also got probably, I'm guessing, I'm not totally sure, but I, I think it has like their 401k contribution information.

Alex Salkever: Or their health information.

Matt DeBergalis: Probably not something-

Alex Salkever: Right. Totally ... that

Matt DeBergalis: people should have.

Alex Salkever: Exactly.

Matt DeBergalis: And, you know, I, I think just across the board, whether we're talking about third-party SaaS that you bought or we're talking about first-party services that you build, the APIs are too coarse-grained. They commingle stuff that has a very different level of, you know, sensitivity or importance.

Matt DeBergalis: Um, that was okay in a world where I had control of the other side and, and knew what that code was doing. But I think we have to look at these agents as [00:15:00] untrusted, maybe even adversarial. How many people are reading this code? Or like think about I think events of the last month with like Anthropic and the DoD really tell us how little we really know about what's happening inside some of the models and the harnesses.

Matt DeBergalis: And if you think about the risk of Data leaks and exfiltration, if you think about this as, like, a savvy adversary, a competitor, maybe even a state actor, this is the biggest opportunity-

Alex Salkever: Yeah ...

Matt DeBergalis: to crack through. I mean, think about what we're doing. We're asking citizen developers to deploy untrusted agents behind the firewall, and we're frantically wiring those things up to every single system we can find.

Matt DeBergalis: And, and every CEO is basically standing up and saying, "We must do this. We must do it now. We have to take more risk. We have to act like a startup." You know, like-

Alex Salkever: Please compromise our APIs ...

Matt DeBergalis: it's, it's a rock and a hard place [00:16:00] because I think, um, the businesses that do this are gonna go so much faster, and they're gonna have such a competitive advantage, but, like, at what cost if you get it wrong?

Alex Salkever: Yeah. No, totally. I mean, like we were, we were-- I was in Amsterdam last week, um, and chatted with, uh, you know, some of the folks on the kernel, on the Linux kernel, and they're seeing now essentially, um, you know, zero-days come in that are out of not modified frontier models like consumer frontier models- Yeah

Alex Salkever: you know, with very simple prompts-

Matt DeBergalis: Yeah ...

Alex Salkever: to call them out. So totally and, like what you're saying, and if s- you really wanna be targeted, it's even worse.

Matt DeBergalis: Yeah.

Alex Salkever: Uh, so what, it, how do you think about solving this? Or, or does like GraphQL start because it is so faceted and can supply-

Matt DeBergalis: Exactly ... an additional layer?

Matt DeBergalis: So, so I'll give you an example. I, I think, I think there's no question that at a high level, AI's gonna change every part of the stack. It's gonna change how we build software. So one example of that is I think we have to move [00:17:00] to call it a zero trust model, where a, a piece of software or an agent has access to the thinnest slice of what it, what it should.

Matt DeBergalis: And GraphQL is a great architectural starting point for that because It has field-by-field semantics. So a GraphQL schema says, "Here are the objects in your business. Each of those objects has a set of properties." We call them fields, right? We can assign, um, rights to those fields. So we can say a given agent, which has, you know, one credential, you get this field, this field, this field, and this field.

Matt DeBergalis: And, um, that agent is gonna make that request explicit. When I call a REST endpoint, I just say, "Here's, here's the arguments. Give me back the result," and the, the API decides what I'm gonna get back. That model doesn't work for an agent because the API, the microservice, doesn't know what the agent's [00:18:00] trying to do.

Matt DeBergalis: It doesn't know what it is, and it doesn't, it doesn't-- It's not in a position to enforce policy

Alex Salkever: And you can't think about what the agent's composing, if the agent's composing together four or five things or, yeah.

Matt DeBergalis: And, and this stuff goes pretty deep. You know, we work with financial companies that, I think one example is support agents only have access to a customer's billing records if there's an open support ticket So those are two totally separate systems that you have to join together, but like that's a law.

Matt DeBergalis: That's, that's not some desired policy. That's actually a legal requirement that's placed on a lot of financial companies as part of a consumer protection stance. So the client has to be able to specify what it's trying to do, and I think GraphQL gives us a template for how to go about doing that, right?

Matt DeBergalis: Because in GraphQL, you write a query, and the query is explicit. It says, "I want, I want X, Y, Z, and W out of the graph, and I want them in this order." And now we're in a position to say, "Yep, that's fine," or like, "If you want that piece of personal information, [00:19:00] we're gonna block that request until we do, uh, an approval process that goes through, you know, somebody who owns the relevant system and can dole out that access."

Alex Salkever: So essentially, almost as a, either as a shim or if they're just using that as sort of a native, uh, you know, primitive construct of their API, it, it allows them to assign much more granularity to access attributes, anything that they wanna share. Yeah. Um, so a question on top of that, 'cause this was actually an interesting real-world, uh, problem that I saw.

Alex Salkever: I think IBM wrote about it or somebody. But, uh, there was an instance where you essentially had two bots, less agents inside of a company. You know, one was tasked with customer, uh, you know, customer mag-- uh, maximizing customer satisfaction. The other was ta- You may have heard this story. Was tasked with, uh, you know, whether to give people refunds or not.

Alex Salkever: And, and like the reward structure was, you know, customer satisfaction was the biggest reward. So it actually started handing out refunds willy-nilly automatically.

Matt DeBergalis: Yeah,

Alex Salkever: yeah. Uh, so like [00:20:00] this is agent to agent. So when we start to have these kinds of dynamics, how, how do, how do you think about writing or at the API level starting to put better strictures or controls around this?

Matt DeBergalis: I think the rabbit hole goes pretty deep here, right? Like you can start at the surface level, which is should an agent be able to give a refund autonomously? Maybe not today, right? Maybe tomorrow, maybe, maybe below 50 bucks, right? Like-

Alex Salkever: Right.

Matt DeBergalis: But you're getting at this much deeper, much more interesting question, I think, which is like what's it really gonna look like to have an autonomous agent that's doing the work that a group of people would've done in the past?

Matt DeBergalis: And how do we, you know, not just think about it in terms of sort of the basics of security, but like maybe a quality metric that comes with that. Um, are we making a good business decision? And how do we measure that? And how do we know? I think the interesting thing is have we really had a good way to do that in the, in the pre-AI [00:21:00] world?

Alex Salkever: Oh, no.

Matt DeBergalis: Absolutely not. Not really.

Alex Salkever: This is a new question-

Matt DeBergalis: Exactly ...

Alex Salkever: and new concept,

Matt DeBergalis: yeah. Yeah. Um, so I think there's gonna be a lot of innovation here, and, um- I mean, it's interesting, like maybe a faster turnaround time on that decision is worth it if you're offering maybe more refunds than you really needed to.

Matt DeBergalis: Maybe it's not. Maybe that trade-off changes over time. And I think we're gonna find a lot of energy around exploring this space of business decision-making with more and more autonomy, because I, I think at the end of the day, it's the clock speed of the business that really determines, uh, that company's prospects.

Matt DeBergalis: I mean, it's a, it's a very rapidly changing world, and if, you know, fast beats right, I think in a lot of, in a lot of moments today, um, and that's an example of it.

Alex Salkever: Yeah, I completely agree 'cause I think that that's like an instance where user expectations, either in the enterprise or consumer, they want everything [00:22:00] faster all the time, you know, and they're...

Alex Salkever: And, and, and it really tends to drive processes in that direction. And sort of the flip side of that, which I'm actually interested in your thoughts on, which I think builds on top of it, is on, on the one hand you have security, on the other hand you have trust, uh, you know, in a two package. And trust will enable speed if you do it right.

Alex Salkever: Uh, so what are... How are you seeing people think, or how do you think about, uh, you know, creating trust systems within these complex autonomous interactions?

Matt DeBergalis: Yeah, and can I add a third? Please. 'Cause I, I think about a stool. There's also, there's experience or friction. 'Cause, 'cause I, I think a great security footprint isn't gonna be worth anything- If you have no customers

Matt DeBergalis: if everybody works around

Alex Salkever: it. If everybody's mad about it. Yeah.

Matt DeBergalis: Yeah. You're right. Yeah. And, and, and they will. So- I think of it as a, it's, it's really a, an experience problem at the end of the day. Like, if we only solve an access control problem, it's putting up a bunch of stop [00:23:00] signs. But if we can look at it as, hey, there is a wealth of incredibly valuable data that you now, you can have access to, and you don't have to ask, because all that is taken care of behind the scenes for you, right?

Matt DeBergalis: And a lot of stuff doesn't require a ton of scrutiny. You probably wanna know what's happening. You probably want an audit log of all your agents and what they're up to. But if we can empower people to take that first step, that builds a lot of momentum, I think, for what comes second and third and fourth.

Matt DeBergalis: 'Cause I, I think everybody had that experience with AI, right? Like, you kinda read about it, but it's when you have your first touch with, oh my gosh, that really went as well as I thought, or you see a hackathon- Yeah ... where somebody, you know, builds something and you're like, "Wait, you built that and you did it in a day?"

Matt DeBergalis: Right? Yeah. I, I- It kind of forces you to rethink your assumptions about

Alex Salkever: a lot of things ... I, I built a custom MD scraper that basically pulls from YouTube, Linked- Yeah ... I mean, like anything I wanted, I'd do the half-hour vibe coding on, on CollectCode. [00:24:00]

Matt DeBergalis: Yeah.

Alex Salkever: I wrote it locally, but still, you know, yeah.

Matt DeBergalis: So I just keep coming back to, I, I think there's a really positive opportunity here for companies to get really better at what they do.

Matt DeBergalis: I mean, I've, I've loved the technology. I've, I've been into infrastructure for a long time, but I think what really drives me for a lot of this is I, I wanna... I'm an end user. I wanna see better products. I wanna see better customer experiences. I want that stuff to really be great. And the AI thing goes way beyond helping companies write that software faster that I use every day.

Matt DeBergalis: It's gonna make the company a more customer-centric company, and it's gonna show up in all kinds of different places, I think. And so it's, it's, um- It is a security and a trust problem, but I think it's also, um, more and more, the way I think about it at least, is you get there by meeting people where they are and showing them a, [00:25:00] an exciting path.

Matt DeBergalis: 'Cause I think everybody wants to be a builder at some level, you know? Like-

Alex Salkever: So- ... it's

Matt DeBergalis: fun

Alex Salkever: So, so related to that, and although much deeper down in the weeds, if you had a wishlist of changes to the MCP spec or to the GraphQL spec to further this vision, what would be the things that you would, you would sort of say you think could really help things along?

Alex Salkever: And if I'm getting you into- How much, how

Matt DeBergalis: much time do we

Alex Salkever: have? If I'm get- if I'm getting, I mean, if I'm getting into political trouble, then you know I-

Matt DeBergalis: No, no, no, no. Um, I mean, I've, I've talked about the GraphQL piece in, in a lot of places. Uh, well, I think the more context you can give the model, the more capable they get.

Matt DeBergalis: So a lot of this stuff, um, in, in both MCP and GraphQL, uh, I think goes to the idea of providing more structured context. There's an opportunity to, I think, provide, uh, a richer amount of, like, if you think about what this means, how does a human think about what something means? Well, they think about [00:26:00] how's it being used elsewhere.

Matt DeBergalis: They think about, um, access patterns. They think about... I mean, all that human judgment, we don't-- It's not always at the front of our mind, but it's what's driving our, our decisions as developers. So we've gotta get all that to the models, and there's a whole bunch of stuff you can do there. I think another piece is just if we think about, um, adapting The APIs that we have, which aren't going anywhere, as you said, to what we're trying to do.

Matt DeBergalis: In practice, an agent in many cases is gonna wanna fetch a bunch of data out of a system, and the only way to do that is gonna be to repeatedly call something to paginate across a whole bunch of stuff. Well, that, that's really not... There's a lot of problems with that access pattern.

Alex Salkever: Sure.

Matt DeBergalis: So I think there are capabilities we can think about.

Matt DeBergalis: SQL's probably a good model for some of this stuff, where the APIs get more sophisticated. We can move beyond sort of a basic request-response model and start talking about things like [00:27:00] aggregation and sorting and filtering, and the kinds of things that an agent can quickly express that can be pushed further down the stack.

Matt DeBergalis: 'Cause the closer we can get those actions down to the core systems, the more efficient all that's gonna get.

Alex Salkever: So, like, essentially the task or the, whatever the, you know, the goal, the work becomes expressed in a more complex query structure that further down it can avoid some of the, some of the pitfalls of traditional API behaviors.

Alex Salkever: 'Cause, '

Matt DeBergalis: cause, I mean, as an example of, like, the higher we can go up in the abstraction chain, the more capable it's gonna become.

Alex Salkever: Sure.

Matt DeBergalis: And, and, you know, I, I think that applies to APIs just as well. And at a high level, I think the question is just, like, do agents want a different kind of set of access patterns?

Matt DeBergalis: Like, most APIs were built either for screens or they were built for least common denominator, you know, service-to-service communication, like a very domain-oriented, you need object 25, here you go. Right.

Alex Salkever: Right? Right,

Matt DeBergalis: right. So I just think what we're gonna [00:28:00] find is that agents want other things than that They're gonna wanna crawl over that data in a different way.

Matt DeBergalis: We can anticipate some of that. I think some of it we'll still learn as we go. But I think there's gonna be a lot of value in adding that kind of higher level semantics to these things. There's a lot more, though. I mean, we could talk about real time and streaming and push. That was a topic from this morning, right?

Matt DeBergalis: Like AI, I think fundamentally is kind of a real time thing.

Alex Salkever: It's certainly accelerating the way we interact with our technology.

Matt DeBergalis: Yeah. And, and I think both MCP and GraphQL have some, you know, room for improvement on that front because again, it wasn't maybe at the heart of what the original workloads most needed.

Matt DeBergalis: Um, there's hints of it, but, but there's a lot more that we can build there.

Alex Salkever: I'm very interested sort of as-- 'cause you live in, at the intersection of two very interesting worlds. I mean, 'cause in GraphQL, I mean, you don't see a lot of people here that are very active in the API spec community as well. Uh, some, but it's, you know- Yeah

Alex Salkever: not, not, you, you sit [00:29:00] r- squarely in the middle and it's, uh, so you mean you have a, a very different and sort of, not different, but a, a nuanced view of this or a, you know, an understanding of both sides that's really different.

Matt DeBergalis: Yeah. It's, it's-- I feel like we've made contact with business world now.

Alex Salkever: Yeah.

Matt DeBergalis: And there's no surprises there at the technical level, but there's a lot...

Matt DeBergalis: Here, here, I'll leave you with this. I've learned over the work we've done that APIs, if you think about it, are a handshake, they're a contract. And in a company, especially a mature company, especially one where you've got thousands of these, this is really a people problem at its heart. This is all about a multi-stakeholder alignment and negotiation.

Matt DeBergalis: I think AI can offer a lot of benefit to navigating those conversations and navigating the kinds of changes that you're gonna wanna make. But I think it's also important to, to remember that, like, none of that's gonna go away. I think we're, we're always gonna have [00:30:00] people accountable for the quality of these systems and the correctness and the reliability and so on.

Matt DeBergalis: And I think we're gonna have people accountable for the quality and correctness and reliability of the agentic workloads that they build. And if we recognize that up front, and if we approach this not just from the point of view of like what's the transport look like and how can we like add these semantics here to this protocol, but we also look at it from the point of view of how is the organization at scale gonna change.

Matt DeBergalis: I know that sounds really abstract, but I've, I've just found over and over and over again, people are, are in a position to help drive these things, but it's really about Because the API is, is, has two sides, or because you might have 1,000 consumers of your API, it really is about an organizational transformation, which is exciting.

Alex Salkever: That's a really good point to end on. I mean, essentially change NCP and plus API equals change management because you've [00:31:00] democratized access to all of these things- At, at the speed of

Matt DeBergalis: AI ...

Alex Salkever: for better or for worse.

Matt DeBergalis: Yeah, exactly. Yeah, it's for better, but-

Alex Salkever: I

Matt DeBergalis: love it ... but it, but it's not going to come without challenges, and I'm just- Right

Matt DeBergalis: I'm really eager to see that through because I, I think, uh, there's a, an exciting world on the other side of it. I,

Alex Salkever: I think somebody said recently, "YOLO responsibly," or something like that, which, uh, but-

Matt DeBergalis: I like that. Yeah.

Alex Salkever: You... Yeah. Well, you will enable- Yeah ... you're, you're enabling parts of that, so. Yeah,

Matt DeBergalis: yeah.

Alex Salkever: Thank you very much, Matt, for taking the time. I appreciate it. Likewise. Um, hopefully get to talk to you at another conference.

Matt DeBergalis: Sounds good.