dbt Coalesce Metadata Shift

The description fields in your dbt project used to be read by nobody. Now an agent reads them.

dbt ships an MCP server now. Model descriptions, column docs, lineage and metric definitions get handed to AI assistants over an API, so a model can pick a table and explain a number.

A column called status with no description used to confuse a new analyst for ten minutes. An agent will guess what it means, answer the CFO with confidence, and nobody sees the guess.

What I’d check:

• How many models have an empty description? In dbt it’s a short script over manifest.json.
• Do metric definitions live in one place, or in a few dashboards with different versions of revenue?
• Do descriptions say what a table excludes? Test accounts and refunds are where agents go wrong.

Teams are already generating those descriptions with an assistant, which is fine for the obvious columns and quietly confident about the ones that matter. A generated line about status still needs someone who knows the business to read it. I’ve watched an assistant pick a deprecated table because it had the nicer name.

What share of your dbt models would pass a “has a real description” check today?