Coalesce Context As Infra

Treat your metric definitions like a Terraform file, or keep the agent away from them.

Every data vendor’s roadmap this year is about context: semantic layers, metadata served over MCP, agents reading your dbt project. I’m less sure most teams are ready for it.

Infrastructure has an owner, a change process and an alert when it breaks. In most of the 50-person companies I see, metric definitions have none of those. Revenue lives in the semantic layer, in two Looker explores and in a spreadsheet the CFO trusts more than either.

An analyst notices when two numbers disagree. An agent picks one and answers.

What that means in practice:

• Every metric an agent can query has a named owner.
• Changing a definition goes through a pull request, with someone from the business on the review.
• A test fails when two sources of the same metric drift apart.

Until those are in place, keep the agent on read-only, low-stakes questions. It’s still useful there.

How many of the three can you tick today? The full checklist I use is here: https://thomasnys.com/data-platform-assessment-checklist/