If context is king, architecture is the castle

**Ryan Donovan** (0:07)
Welcome to The Stack Overflow Podcast, a place to talk all things software and technology. I'm your host, Ryan Donovan, and I'm recording live at the AI Agent Conference in New York. And I'm here talking context with Matt DeBerglis, CEO of Apollo GraphQL. So welcome back to the show, Matt.

**Matt DeBerglis** (0:25)
Thanks. It's good to talk to you again.

**Ryan Donovan** (0:27)
Yeah.
Everybody seems like they're talking about context now. Talk to a few folks at this conference. And everybody who's doing context has some sort of graph, knowledge graphs. How are you thinking about context?

**Matt DeBerglis** (0:41)
Well, we've been championing architecture around the industry standard way of talking to APIs like a graph. It's called GraphQL. GraphQL got very popular, especially in larger companies where you've got hundreds or thousands of different systems. They were built in different eras with different technologies. And a lot of the early adopters of GraphQL were trying to solve the customer experience problem. Like I'm a hotel and I want to offer people a way to not just reserve a room, but also get a restaurant reservation or late checkout as part of that. So I got to glue a bunch of systems together in a new way to meet the customer. And by the way, that customer wants to unlock their phone or unlock their room with their phone. So the heart of that solution was the idea that you don't want to look at your APIs as these deeply technical request response things. You want to think about the objects that those APIs represent and offer. And if you do that, you can build really great software that has a consistent experience and so on. Guess what agents need, right? Agents need semantics. They need more than just the wires to go talk to everything. They actually need a lot more on top of that. They need to know how something is meant to be used, how it relates to something else.
They need to have all this metadata that comes with that. They need structure so you can do closed loop development, strong typing and things of that sort.
We're very excited about solving those problems. We think there's a need for a principled architecture, and we think GraphQL is the one that has a great track record and makes a lot of sense for this.

**Ryan Donovan** (2:23)
And when I talk to folks about, you know, knowledge graphs, they usually comment as like a database, and you have it as a sort of API level. Do you think there's the sort of context, knowledge graph is reducing the distance between an API and a database?

**Matt DeBerglis** (2:39)
I mean, agents are going to need to talk to anything and everything. I think we've all had that experience in different ways, right? There's just the world where it doesn't have access to it and the world where it does. So the question to me is, what's out there? And what do the agents have to talk to?
I mean, one example is, if you want your agent to do something, that's an API call at the end of the day. But there's also a lot of examples where if you think about the meaning of a piece of data, it's a lot more than individual systems. It's the business logic and it's all of the complexity and value that comes with that. So if I want to have my agent know, I'll go back to the hotel example, what rooms are available? That's an API call. How much is it going to cost to check out at 4 p.m. instead of noon? That's an API call.
And those are your own systems. Think about third-party systems, same story, right? If we do this at our own company, if I want to know who's interviewing and what stage are they at, or who are we selling to and what stage are they at? Those are API calls into SaaS products.

**Ryan Donovan** (3:47)
There's almost like a semantic layer already built into the API calls itself, right?

**Matt DeBerglis** (3:52)
It's lurking inside. But in the past, to get the semantics of what's inside your APIs, you used human developers that made judgment calls and interpreted those APIs as part of the work they do to build software.
Now, we want a world where agents can do that, and if an agent's going to make a good decision about that, it needs more of that context and more of that intent that comes along that was always implicit before. And that's why a layer between the agents and the APIs is so essential to get that right.

**Ryan Donovan** (4:27)