An AI Agent That Builds and Tests dbt Models? We Built One to Find Out.

What happens when you give an AI the tools to not just suggest code, but to actively build and validate data models? To find out, I connected Google's AI frameworks with dbt's new Fusion engine. The result is a functional, if experimental, agent that interacts with a dbt project. This isn't a polished product. It's a hands-on test of how autonomous AI could function in analytics engineering.

My background is in software engineering, but my work had drifted from daily coding. The arrival of capable AI tools rekindled that direct, creative engagement with building systems. As I moved into data, I wanted to recapture that energy. The recent release of dbt's Fusion engine—with its real-time parsing and deterministic compilation—provided the missing piece. It allows for immediate validation of SQL and YAML against the project graph. This means an AI's output can be checked for correctness instantly, turning potential mistakes into fast feedback.

I combined several technologies to make this work. Google's Gemini models handle the reasoning. The dbt MCP server safely exposes dbt's capabilities—like listing models or running commands—as tools the AI can use. Google's Agent Development Kit (ADK) provides the framework to orchestrate the agent's behavior. With dbt Fusion as the foundation, these components create a system where the AI can propose a model, compile it, analyze the logs for errors, and revise its approach.

In practice, I built an agent with specialized components. One tool handles local dbt compilation via Fusion, validating syntax and dependencies before any warehouse execution. Another connects to the dbt MCP server for project metadata and lineage. A separate subagent was designed to analyze modeling logic and best practices. This structure allows the system to move beyond simple code generation into a iterative development loop.

The experiment suggests a shift. Instead of acting as an advanced autocomplete, AI configured this way begins to operate like a junior team member: it can take instruction, use professional tools, check its work, and learn from errors. The guardrails provided by dbt Fusion and MCP make this exploration feel substantive rather than speculative. For teams investing in modern data stacks, this integration points toward a future where AI assists not just in writing code, but in the entire lifecycle of data product development.

Source: dbt Labs Blog