There’s been a lot of excitement lately around tool protocols like MCP for AI agents. They make it easier for models to discover and call external tools.

But when you look at real systems, agents still need a structured way to interact with application backends - services, data, and business logic.

The post argues that MCP solves the tool interface problem, but not the application architecture problem.

There seems to be a missing layer between agents and backend systems: a graph of APIs and data that agents can query and compose. Technologies like GraphQL can potentially serve as that layer by exposing a typed, discoverable interface to the application backend.

Curious how others are thinking about this layer when building real agent systems.