dbt Wizard AI agent shows data impact as a DAG, not a diff

Why a dedicated agent for dbt projects

In the session, Sam Ferguson, a product designer on dbt Wizard, and Ryan Eakman, senior staff software engineer at Fivetran, argued that analytics engineering needs its own AI agent rather than a general-purpose coding agent pointed at a dbt project.

Ferguson's starting point was a shift in the work. As agents take over more of the execution of data work, the human job moves from writing code to judging it. According to the session, generic coding agents were not designed to make that judgment easy. A diff shows which lines changed and a chat log shows the conversation, but neither shows how the underlying data changed or whether the result is right. Teams end up reconstructing that context by hand each time an agent finishes a task.

dbt Wizard treats a dbt project as one connected system rather than a folder of files. It validates every change against a development schema before anything touches production, and it presents the impact as a DAG so a reviewer can see what moved downstream.

Design choices and the reported results

Ferguson named several deliberate choices behind the product:

• 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, so it does not have to search and guess.
• Validation and visual review built into the workflow instead of added afterward.

According to Ferguson, design partners using Wizard have reported a 15 to 20 percent reduction in production incidents since adopting it. This figure comes from the presenter and was not accompanied by methodology in the source.

dbt Wizard is available today across three surfaces: embedded in dbt Studio and as a standalone workspace inside the dbt platform, from the command line, and through the new desktop app.

Live demo of the desktop app

Eakman demonstrated the desktop app using two fictional dbt projects, Jaffle Games and Jaffle Shop. The scenario had the business side and the game design team defining "active player" differently. He showed Wizard forking a conversation to explore a second question without losing context on the first, then working through both in parallel.

Each time Wizard made a change, it surfaced the assumptions behind it. In one case it flagged that a crash metric relied on null session durations even though an unused crash flag already existed in the data. The session described this as a bug the agent caught rather than introduced.

The validation flow separates cheap operations, such as compiling and parsing code, from expensive ones, such as materializing data. Validation runs through an isolated sub-agent, so the check does not inherit the assumptions of the agent that made the change. After validation passed, Wizard could open a pull request directly, carrying the assumptions and context from the conversation into the PR description.

How it compares with Claude Code, and availability

In the Q&A, Eakman was asked how Wizard differs from Claude Code for dbt-specific work. He answered that Wizard ships with dbt-specific tools built in, so a team does not need to assemble its own MCP servers and skills. A general-purpose tool, he said, has to serve every use case, which limits how far it can steer an agent toward dbt-specific patterns.

He also confirmed that Wizard supports skills and MCP connections through its underlying harness, and that it automatically isolates schemas per work tree so parallel agents do not overwrite the same database objects.

For teams already using dbt, the pitch is a review process centered on data impact rather than line-level diffs. dbt Wizard is available now through dbt's product page, with setup details in the Wizard documentation and the Wizard CLI guide. The desktop app is in private beta, and sign-up is open.