GraphQL meets the agent era: Matt Debergalis on APIs, MCP, and enterprise AI

Apollo GraphQL brought GraphQL to the enterprise: a query language for APIs that lets you compose data from multiple disparate systems with a single query. But as AI agents start talking to APIs on behalf of users, the question shifts from "how do I fetch data?" to "how does an agent understand what this data means?"

At HumanX 2026, WorkOS CEO Michael Grinich sat down with Apollo co-founder and CTO Matt DeBergalis to talk about why GraphQL's schema and type system make it interesting for the agent era, and what enterprise AI adoption actually looks like on the ground.

Agents need semantics, not just endpoints

The core insight DeBergalis shared is straightforward: agents connecting to APIs need more than endpoint URLs. They need to understand what those APIs mean, how objects relate to each other, and how they're intended to be used. That's what GraphQL's schema and type system provide—a structured, typed layer over your API surface that encodes object relationships and field-level descriptions.

"The more information you can feed an agent about what an API really means and how it relates to other objects in your business, the more interesting it gets," DeBergalis said.

GraphQL was built to let a single query reach across multiple backend systems and return exactly the shape of data you need. That composability is well-suited to what agents need to reason about complex enterprise environments.

GraphQL And mcp: complementary, not competing

Fresh from the MCP Summit in New York, DeBergalis sees strong complementarity between GraphQL and Anthropic's Model Context Protocol. The distinction is clean: MCP defines how agents connect to systems—the transport, communication, and discovery mechanism. GraphQL provides the typed schema describing what those systems contain and how their data relates, which agents can use to construct valid queries.

Apollo already ships a product that lets you define MCP tools in terms of GraphQL queries, bridging the two worlds directly.

"A lot of the questions about how to expose what you've got—especially 25-year-old systems you're not about to rewrite—is about presenting it in a form agents can consume," DeBergalis said.

This matters for any enterprise sitting on legacy infrastructure. You don't need to rewrite your backends. You need a typed schema layer that makes them legible to agents.

Enterprise modernization is about people now, not just technology

DeBergalis shared a sharp observation about how AI adoption is playing out inside enterprises. GraphQL almost always shows up as part of a modernization effort—moving to the cloud, rebuilding APIs, consolidating microservices. But this wave of modernization isn't purely technical. AI is changing how people work together, reshaping workflows, and altering org charts.

"Every conversation I have—the first question is about MCP or agent discovery. The second question is always about the people," DeBergalis said. "What are other companies doing? How do I handle this change?"

The technology questions have clear answers. The organizational questions are harder—and they're the ones that determine whether adoption actually sticks.

The treasure trove inside your systems

For DeBergalis personally, the biggest breakthrough has been using Claude Code to query internal systems—customer conversation transcripts, operational data—all at the touch of a button. He sees this as the beginning of a much larger story: enterprises and startups alike are sitting on value trapped inside SaaS tools and microservices that were never built for AI-era access.

The data is there. The systems are there. What's been missing is a way to make all of it available to agents that can actually reason over it. A typed, self-describing API layer like GraphQL helps fill that gap—giving agents a schema they can introspect to understand available data and relationships without hardcoded knowledge of each backend.

Non-developers are building agents

One of the most telling anecdotes from the conversation: Apollo's head of talent—not a developer—has been building an agent that reviews interview notes and provides coaching to hiring managers.

"It's the sort of thing we never would have done as thoroughly or quickly," DeBergalis said. "Now it's just part of our standard."

The challenge is enablement. Demos spark hunger, but the details of helping people build on what they've seen are where things get hard. DeBergalis noted that the traditional "learning to program" curriculum isn't quite right anymore. What people need now is a sense of systems design and taste—understanding how to decompose problems, connect services, and evaluate tradeoffs—not React or Python syntax.

The golden age of building

DeBergalis is also seeing AI change procurement decisions in real time. Apollo has started not renewing certain SaaS subscriptions—not to cut costs, but because teams realized they could build tailored replacements quickly.

"That visceral sense of joy and relief—especially for folks that spend their day operating stuff—it strikes a nerve that accelerates how quickly people want to march down this path," he said.

When building becomes fast enough that it's easier than buying, procurement logic shifts. That's the moment DeBergalis thinks we're in right now.

This interview was recorded at HumanX 2026 in San Francisco.