Top AI Product

Building a machine-learning pipeline is still slow, manual work — feature engineering, model selection, tuning, glue code. Snowflake’s new Data Science Agent, unveiled at its 2026 Summit, tries to collapse that into a prompt: describe the problem in plain English and it returns a fully executable ML pipeline.

## What it does

The agent uses multistep planning to break a problem into distinct stages, then iterates and adjusts as it goes, choosing the best-performing tools for each step rather than running a fixed template. Anthropic’s reasoning models run under the hood. The output isn’t a suggestion or a notebook stub — it’s an executable pipeline that lives where the data already is, inside Snowflake.

## Part of an agentic push

It arrives alongside a broader Snowflake ML refresh: distributed training APIs, native experiment tracking, ML Jobs for orchestrating pipelines, a Feature Store for low-latency predictions, and built-in observability. The throughline is Snowflake’s stated goal of becoming the “control plane for agentic AI” — keeping the agents, the models, and the governed enterprise data in one place instead of shuttling data out to a separate ML stack.