K-Dense-AI/agentic-data-scientist

What it solves

Agentic Data Scientist provides an automated framework for tackling complex data science tasks that typically require iterative planning, coding, and validation. It eliminates the need for manual trial-and-error by separating the planning phase from the execution phase and implementing a self-correcting loop to ensure results meet specific success criteria.

How it works

The project uses a multi-agent workflow built on Google's Agent Development Kit (ADK) and the Claude Agent SDK. It operates in two primary modes:

1. Orchestrated Mode: A full multi-agent pipeline where a Plan Maker and Plan Reviewer create a validated analysis plan. A Stage Orchestrator then manages a Coding Agent (powered by Claude Code) and a Review Agent to implement the plan stage-by-stage. A Criteria Checker and Stage Reflector track progress and adapt the plan based on discoveries during execution, ending with a Summary Agent producing a final report.
2. Simple Mode: A direct coding approach for quick scripts or question answering without the planning overhead.

The system integrates with Model Context Protocol (MCP) servers for live documentation and utilizes a library of 143 scientific skills for specialized data analysis.

Who it’s for

Data scientists and engineers who need to automate complex analysis pipelines, perform scientific research (such as differential expression analysis), or generate publication-ready reports from raw datasets.

Highlights

• Adaptive Workflow: Iteratively refines plans and implementations based on real-time results.
• Scientific Integration: Access to 143 specialized scientific skills and databases (e.g., UniProt, PubMed).
• Continuous Validation: Tracks objective success criteria at every stage to prevent wasted effort.
• MCP Support: Uses the Context7 MCP server for up-to-date library and framework documentation.