Google AI Introduces DS STAR: A Multi Agent Data Science System That Plans, Codes And Verifies End To End Analytics

Abstraction: Multi-agent system converting natural language to Python for heterogeneous data analytics

Key points:

• DS STAR (Data Science Agent via Iterative Planning and Verification) reframes data science as Text-to-Python over mixed file types (CSV, JSON, Markdown, text) rather than Text-to-SQL
• Six specialized agents: Aanalyzer, Aplanner, Acoder, Averifier, Arouter, Afinalyzer — plus Adebugger and a Retriever for robustness
• With Gemini 2.5 Pro as the base LLM and up to 20 refinement rounds, DS STAR raises DABStep hard-task accuracy from 12.70% to 45.24%
• On KramaBench, DS STAR scores 44.69 vs. 39.79 for the best alternative (DA Agent); retrieval uses Gemini Embedding 001 to select top-100 relevant files
• Ablations confirm analyzer descriptions and routing are critical; architecture is model-agnostic (tested with GPT-5)

Connections: Google · Gemini · AI Agents · Agentic Coding · Large Language Models

Source: https://www.marktechpost.com/2025/11/06/google-ai-introduces-ds-star-a-multi-agent-data-science-system-that-plans-codes-and-verifies-end-to-end-analytics/