This summer, dbt Labs released dbt Wizard, a coding agent specifically designed for working on dbt projects.
Overview
What is the dbt Wizard? dbt’s website describes it as:
“Built for governed data development in dbt, dbt Wizard understands your project, routes to the right dbt tools, and validates work with awareness of warehouse operations. Use it to investigate failed runs, debug models, assess impact, make changes, and ship trusted data work in one place.”
The dbt Labs team has been working on dbt for 10 years now. They have not only the best expertise in managing performant dbt projects, but also a bunch of data from their professional services and from managing dbt Cloud about the mistakes users make.
They already released agent skills, and a custom agent was a logical next step. These, of course, build on top of the dbt Copilot that struggled to gain mass adoption.
With the Wizard agent, you get the guidance built-in of an agent that knows best practices and workflows for analytics engineering work in dbt.
And in a win for users of dbt Core, you can use dbt Wizard in dbt Cloud (need an account) or on the terminal (no account needed).
Our first impression
We tested out dbt Wizard shortly after release.
Installation and setup were super easy with an existing dbt project. It was able to find my profiles.yml and plugging in an Anthropic API key worked the first try.
We gave it a small task, to rename a column in a way that would cause breaking changes downstream.
dbt Wizard completed it easily and quickly. Overall, a very smooth demo.
dbt Wizard vs. Claude Code
This boils down to a question of whether you should rely on agents that are trained to perform really well for engineering problems like Claude Code, or something fine-tuned for a specific use case like dbt Wizard.
For Claude Code specifically, it’s hard to ignore that the Claude Code team has shipped so much vibe code that Claude Code actually ranked middling on the terminal-bench harness leaderboard (the leaderboard has since removed 3rd-party harnesses).
On the other hand, Anthropic and OpenAI heavily subsidize tokens spent through Claude Code and Codex. dbt Wizard is bring-your-own-key (BYOK) for most of the big model providers, which means you can hook up your existing account and pay API prices.
That was enough of a deterrent for me to stick with Claude Code if I’m getting good enough productivity from my existing dbt agentic engineering setup. The good news is that switching costs are very low if we change our mind in the future.
If we were getting started from scratch today and wanted to trade token spend for speed, I’d consider the dbt Wizard agent in the data stack design process.