Curious what people are seeing with domain-grounded coding agents vs. general-purpose ones (Copilot, Claude Code, etc.) for data/ML work specifically. The pitch from the vertical tools (Databricks' Genie Code is the one I've used) is that grounding in your actual schema/lineage/governance layer beats a general agent guessing from context alone. Databricks claims a jump from ~32% to ~77% success rate on real data science tasks after adding that grounding. Haven't independently verified that num...