dbt Wizard: An AI Agent That Shows How Data Actually Changes

At a dbt Summit 2026 breakout session, product designer Sam Ferguson and Fivetran senior staff software engineer Ryan Eakman argued that analytics engineering deserves its own AI agent rather than a general-purpose coding tool pointed at a dbt project.

Ferguson explained the gap dbt Wizard was built to close. As agents handle more execution, humans shift from writing code to judging it, but generic agents don't make that judgment easy. A diff shows changed lines and a chat log shows the conversation, yet neither reveals how the underlying data changed or whether the result is correct.

dbt Wizard treats a dbt project as one connected system and validates every change against a development schema before production. It presents impact as a DAG instead of a diff, letting reviewers see what moved downstream. Ferguson said design partners have reported a 15 to 20 percent drop in production incidents since adopting it.

The agent is available in dbt Studio and as a standalone workspace in the dbt platform, from the command line, and through a new desktop app now in private beta.

Eakman's live demo used two fictional projects, Jaffle Games and Jaffle Shop. In a scenario where business and game design teams defined "active player" differently, Wizard forked a conversation to explore a second question while keeping context on the first. It surfaced the assumptions behind each change and caught a crash metric relying on null session durations despite an unused crash flag in the data.

Validation separates cheap operations like compiling from expensive ones like materializing data, running checks through an isolated sub-agent. Once validation passes, Wizard can open a pull request carrying the conversation's context into the description.

Asked how it differs from Claude Code, Eakman said Wizard ships with dbt-specific tools built in rather than requiring teams to assemble their own MCP servers and skills, and that general-purpose tools must serve every use case, limiting how far they can steer toward dbt patterns. He confirmed Wizard supports skills and MCP connections and isolates schemas per work tree so parallel agents don't collide.

Source: dbt Labs Blog