Fail Before You Run: Typed Data Contracts for AI-Written Pipelines
Your coding agent writes a 300-line pipeline. It runs for forty minutes, then fails because a column it assumed was an integer is a string. It reads the stack trace, guesses, and runs it again. Or worse: the run succeeds, a column has been quietly renamed, and a dashboard three steps downstream breaks on Monday morning.
These are not model problems. They are interface problems. Most data stacks leave the shape of the data implicit: which types each step expects, which columns downstream tables depend on, what a model is supposed to return. A person learns those assumptions over months. An agent only knows what is in its context and what your errors tell it, so every hidden assumption becomes another round of trial and error, paid for in tokens, compute, and the time of whoever reviews the result.
Bauplan 0.3 makes those assumptions explicit and checkable. Every model declares the typed schema of what it reads and what it returns, in the same code that produces it. Before a job starts, Bauplan checks that each step's inputs match what upstream produces, so a mismatch fails at planning time, not forty minutes into the run. At runtime, it verifies that each model returns exactly what it promised, so a wrong type or a renamed column becomes a precise error the agent can fix, not a bad table in your lakehouse. The same contracts work across Python and SQL, so a mixed-language pipeline is checked end to end.
In this webinar, Ciro Greco and Giacomo Piccinini show what this looks like on a real pipeline. A coding agent builds a new model from a contract, gets caught before anything runs, fixes its own mistake on a branch, and merges only when the data matches the spec. Along the way they share the design principles behind 0.3, including the ones you can apply to any data stack.
What you'll learn
Common ways agents fail when writing data pipelines, and which of them better interfaces can prevent
How to catch type and schema errors before you pay for compute
Contract-first development: write the spec, let the agent write the pipeline