Data science by committee — literally

What it does

Agentic Data Scientist is a Python CLI that orchestrates multiple LLM agents — planners, reviewers, coders, and validators — to carry out data science tasks end-to-end. You feed it a query and optionally some CSVs; it spins up a workflow where one agent drafts a plan, another critiques it, a third writes code via Claude Code, and others check whether the job actually got done. Results land in a local working directory.

The interesting bit

The framework explicitly separates planning from execution and bakes in iterative refinement loops — plan, review, parse, code, review again, check criteria, reflect, repeat. This isn’t just prompt chaining; it’s a structured bureaucracy of models, with the “Stage Reflector” agent empowered to replan mid-flight when the data surprises everyone. The README’s ASCII workflow diagram is endearingly earnest about this.

Key highlights

• Built on Google’s ADK and Anthropic’s Claude Agent SDK, with planning/review routed through OpenRouter and coding locked to Anthropic’s API
• Two modes: orchestrated (full multi-agent theater) and simple (direct coding, no planning overhead)
• MCP server integration, including Context7 for live library documentation lookup
• Network access toggle — agents can search the web unless you explicitly cage them
• Output persists to ./agentic_output/ by default; --temp-dir for throwaway runs

Caveats

• Requires two separate paid API keys (OpenRouter + Anthropic) before you can even test it
• The “Claude Scientific Skills Integration” is mentioned but not explained in any detail — unclear what this concretely adds
• No discussion of cost controls, token budgets, or rate-limiting for the multi-agent loops, which could get expensive fast on complex tasks

Verdict

Worth a look if you’re already paying for multiple LLM APIs and want a structured, auditable alternative to vibe-coding your analyses. Skip it if you need predictable costs, offline operation, or a single-vendor stack — this is fundamentally a cloud-dependent orchestration layer.