When skills meet scale: an agent assembly line for dbt

As data teams explore AI, many are looking to move beyond simple text generation toward practical, ""agentic"" workflows. By leveraging dbt, data practitioners can support safe, multi-agent orchestration to modernize the analytics engineering SDLC.

Using dbt agent skills (now a part of dbt Wizard) and the dbt Fusion engine, my team has been exploring a development workflow where AI agents securely read metadata for automated discovery, translate complex legacy logic into dbt models, and generate PRs for code review. This approach relies heavily on a read-only guardrail architecture that ensures humans remain in control at every step.

In this session, we'll dive into this problem space and explore a practical architectural approach. We will discuss how to scope agent permissions to read-only, what these guardrails look like in practice, and where human review remains non-negotiable. You'll leave with a concrete blueprint for deploying multi-agent dbt workflows safely, along with a clear picture of what ""safe"" actually requires in a production data SDLC.