Inside dbt Wizard, the AI agent built to show its work

In a dbt Summit 2026 breakout session, Sam Ferguson, a product designer on dbt Wizard, and Ryan Eakman, senior staff software engineer at Fivetran, made the case for why analytics engineering needs an AI agent of its own, rather than a general-purpose coding agent pointed at a dbt project.

Ferguson opened with the problem dbt Wizard was built to solve. As agents take on more of the actual execution of data work, the job for the humans involved shifts from writing code to judging it, and generic coding agents were not designed to make that judgment easy. A diff shows which lines changed.

A chat log shows the conversation. Neither shows how the underlying data actually changed, or whether the result is right, which leaves teams reconstructing that context by hand every time an agent finishes a task.

dbt Wizard's answer is to treat a dbt project as one connected system rather than a folder of files, and to validate every change against a development schema before it ever touches production. When dbt Wizard makes a change, it shows the impact as a DAG rather than a diff, so a reviewer can see what actually moved downstream. According to Ferguson, design partners using Wizard have reported a 15 to 20 percent reduction in production incidents since adopting it.

Underneath that pitch, Ferguson pointed to a handful of deliberate design choices: a harness built specifically for analytics engineering rather than a general chat interface, a native metadata engine that gives the agent structured access to a project instead of making it search and guess, and validation and visual review built into the workflow itself rather than bolted on afterward. dbt Wizard is available across three surfaces today: embedded in dbt Studio and as a standalone workspace inside the dbt platform, from the command line, and now through a new desktop app.

Eakman then ran a live demo of the desktop app using two fictional dbt projects, Jaffle Games and the familiar Jaffle Shop. Working through a real scenario, where the business side and the game design team defined "active player" differently, he showed dbt Wizard forking a conversation to explore a second question without losing context on the first, then working through both in parallel. Every time dbt Wizard made a change, it surfaced the assumptions behind that change, including one case where it flagged that a crash metric relied on null session durations even though an unused crash flag already existed in the data, a bug the agent caught rather than introduced.

The demo also showed Wizard's validation flow splitting cheap operations, like compiling and parsing code, from expensive ones, like materializing data, and running validation through an isolated sub-agent so the check wouldn't inherit the same assumptions as the agent that made the change. Once validation passed, Wizard could open a pull request directly, carrying the assumptions and context from the conversation into the PR description.

In the Q&A, Eakman fielded a question about how Wizard differs from Claude Code for dbt-specific work, and answered that Wizard comes with dbt-specific tools built in rather than requiring a team to assemble its own MCP servers and skills, and that a general-purpose tool has to serve every use case, which limits how far it can steer an agent toward dbt-specific patterns. He also confirmed Wizard supports skills and MCP connections through its underlying harness, and that it automatically isolates schemas per work tree so parallel agents don't clobber the same database objects.

dbt Wizard is available today through dbt's product page, with setup details in the Wizard documentation and the Wizard CLI guide. The desktop app is in private beta with sign-up open now.

Watch every keynote and breakout from dbt Summit 2026 on demand. Watch dbt Summit 2026 on demand →