You want to give AI access to your services, so you put an MCP server on top of them.

You pick the operations one agent needs, write a tool for each, write the descriptions, parameter docs, and auth rules. Then you join data across two services, so you write a tool for that too. You just built a backend for frontend, for a consumer who can’t tell you what it needs, till it needs it.

In this talk we go through the evolution of MCP tools. From putting an MCP facade on top of your REST API, through to purpose built capabilities tailored for. How to account for toolset management, ACL, cost estimates, development speed. When the tool list becomes unmanageable for an LLM, you use a search and call tool.

But what happens when you can represent the capabilities of your organisation as a graph? What does that unlock for your agents?

A graph is one contract across the services you already have, in a form a machine can read. Nobody hand-picks what goes in it. It is everything your company can do, in one place.

That changes what an agent can ask for. Today the best answer is search tool and call tool. The next generation is prompt and execute. The unit stops being a tool someone wrote in advance, and becomes a capability your graph can satisfy on demand.