A data engineering agent in practice

The most impressive thing an experimental BigQuery data engineering agent did was undo its own mistake. From plain-language prompts, it built a loading pipeline from Cloud Storage, enriched a schema using built-in AI functions, generated data-quality checks and wrote table metadata to support conversational analytics.

Those are the foundational, repetitive tasks that eat a large share of a data engineer's week. Automating them frees skilled people for work that needs judgment: modeling the business correctly, deciding what quality means for a given dataset, and working with the teams who consume the data.

But the moment that stuck with me was when the agent detected that one of its own changes was wrong and reverted it. That is the difference between a code generator and a collaborator, and Google's introduction to the agent shows how far they intend to take it. In data work especially, a wrong schema change or a silent load error can spread into dozens of downstream reports before anyone notices, so an agent that checks itself is worth far more than one that is merely fast. For regulated enterprises, the generated metadata and quality checks may end up more valuable than the pipeline itself, since they are the evidence governance teams need.

Judge a data agent by how it handles its own errors, not by how fast it writes SQL.