#sub-agent

Sep 3, 2026

A Data Agent Should Be a Sub-Agent

Here’s how most data agents get built. Someone connects one to a semantic model, opens the big chat pane in the middle of the screen, and types “how are sales?” A number comes back. Everyone nods. Nobody opens it again. That’s what most people think a data agent is for: another chat interface. One more box to type into, sitting next to Copilot, ChatGPT, Claude, and whatever your company bought last quarter. I think a data agent should be a sub-agent. It’s a narrow specialist with one job written into its instructions, and a bigger agent calls it when a question needs it. People talk to Copilot, inside the tools they already work in. Copilot calls the data agent. Once you look at it that way, two things change. You stop building one data agent per semantic model, and you start writing instructions per purpose.