Observation
The slowest part of agentic EDA is defining the purpose of the analysis.
What changed
Agents make the mechanical parts cheaper. They can summarize columns, flag nulls, cluster edge cases, and produce first-pass questions very quickly.
What did not change
They still need a human holding the frame. If the question is weak, we miss the answer by a wider margin. If we're relying on correlation rather than causation, then the signal is even more fragile, inflated, and likely incorrect.
Current rule
- Start with one operational question.
- Define one success condition before the run starts.
- Cut any branch that does not change a product or data decision.