Agentic Scientist
Sakana AI's open-source pipeline that turns coding agents such as Claude Code, Codex, Cursor and Gemini CLI into AI scientists: from a task file they propose ideas, run and score experiments and ablations, then write the paper.
Updated
✨ Try AgentScout's free AI image generator →Good for: writing code, coding in the terminal, research
Use cases
About Agentic Scientist
Agentic Scientist is listed in our research category and is aimed at builders. Sakana AI's open-source pipeline that turns coding agents such as Claude Code, Codex, Cursor and Gemini CLI into AI scientists: from a task file they propose ideas, run and score experiments and ablations, then write the paper. Pricing, as the maker states it: Free and open source (Apache-2.0); you pay for the agent and model API keys or subscriptions you use. Maturity: Active. We last checked the official link on 2026-10-20.
What is Agentic Scientist?
Agentic Scientist, from Sakana AI (the team behind The AI Scientist), treats coding agents as AI scientists. You give it a task.md describing a research problem; the agents propose ideas, run experiments, score them against the task rubric and iterate toward stronger programs. The pipeline then runs baselines and ablations, a sealed final evaluation and post-hoc analyses, and writes, compiles and audits a paper against that evidence, with every stage kept under one run directory.
It drives Claude Code, Codex, Cursor, OpenCode, Gemini CLI and pi through the Headless CLI, so the same experiment can be repeated across agents and models. Every phase runs in its own Docker container by default, because the agents run with full tool permissions, and a browser-based Run Inspector lets you review each run. The accompanying paper is on arXiv (2610.18933).
Sources: the official README and our directory listing, checked 2026-10-20. Official site: github.com/SakanaAI/AI-Scientist-v3.
Who is Agentic Scientist for?
Researchers and ML engineers who want to hand a well-defined research task to coding agents and get back experiments, ablations and a draft paper they can check, and people comparing how different agents and models do on the same task.
Agentic Scientist at a glance
- Open source: Yes (Apache License 2.0)
- GitHub: SakanaAI/AI-Scientist-v3 · 428 stars · 65 forks (as of Oct 20, 2026)
Key features
- Six phases: task drafting, investigation, ablations, sealed final evaluation, post-hoc analysis and paper
- Works with Claude Code, Codex, Cursor, OpenCode, Gemini CLI and pi (agent.name=...)
- Each phase runs in its own Docker container for safety and reproducibility
- Optional external AI review and revision phases for the paper
- Run Inspector web UI to browse ideas, attempts, results and the manuscript
- CPU-only example task (circle packing) plus GPU task presets
Agentic Scientist FAQ
What is Agentic Scientist?
Agentic Scientist, from Sakana AI (the team behind The AI Scientist), treats coding agents as AI scientists. You give it a task.md describing a research problem; the agents propose ideas, run experiments, score them against the task rubric and iterate toward stronger programs. The pipeline then runs baselines and ablations, a sealed final evaluation and post-hoc analyses, and writes, compiles and audits a paper against that evidence, with every stage kept under one run directory.
Who is Agentic Scientist for?
Researchers and ML engineers who want to hand a well-defined research task to coding agents and get back experiments, ablations and a draft paper they can check, and people comparing how different agents and models do on the same task.
Is Agentic Scientist free?
Agentic Scientist's pricing, as the maker states it: Free and open source (Apache-2.0); you pay for the agent and model API keys or subscriptions you use. Check the official site for current plans.
Is Agentic Scientist open source?
Yes. Agentic Scientist is released under the Apache License 2.0, according to its GitHub repository.
What do I need to run Agentic Scientist?
Docker, Python 3.11+ with uv, and API keys or a subscription for the coding agent you pick (for example a Claude login for the default claude agent). GPUs are only needed for the GPU task presets; the circle-packing example runs on CPU.
Is it safe to run?
The project warns that it launches coding agents with full tool permissions, so they can run arbitrary commands. It recommends keeping runs inside its Docker image, only running tasks and sources you trust, and reviewing run directories before sharing them.
What are alternatives to Agentic Scientist?
Similar tools include EvoScientist, OpenScience, Hyperresearch, ToolUniverse, Firecrawl and Bruin.
Agentic Scientist ratings and reviews
Agentic Scientist alternatives
Other research tools in our directory. See all research tools →
EvoScientist
Self-evolving AI scientist system that autonomously explores research ideas, runs experiments and improves with growing skills and memory.
OpenScience
Research agent workbench that reads the literature, writes and runs code and experiments and logs every step, with skills and connectors for databases like PubMed, UniProt and arXiv.
Hyperresearch
Deep research harness that turns Claude Code or Codex into a research agent building a persistent, searchable wiki, with a hosted API and MCP.
ToolUniverse
Harvard's collection of scientific tools for AI scientist agents.
Firecrawl
Turns websites into clean, LLM-ready data for agents.
Bruin
Open-source, AI-powered data platform covering ingestion, pipelines and quality checks.
Related lists
🏆 See also: Agent-Reach alternatives · All best-of lists
Badge for makers
Built Agentic Scientist? Add this badge to your README or site. It links back to this page. About badges
