The researchers proposed an architecture called TabFM Auto, which keeps TabFM frozen as the inference engine while an LLM-driven agent takes over pipeline tasks such as inspecting column names, writing data transformations, selecting in-context rows, and calibrating predictions.