Creativity · Agent Protocol
Sakana AI Scientist: Fully Automated Research Pipeline
The AI Scientist, published by Sakana AI (Tokyo) in August 2024 and evolved through v2 in 2025, is a fully-automated research pipeline: it ideates novel ML research directions, writes experimental code, runs experiments on a GPU, analyzes outputs, and drafts a LaTeX paper complete with references and related-work discussion. It's the first compelling public demonstration that a closed-loop agent can produce end-to-end ML research — imperfect but real.
Protocol facts
- Sponsor
- Sakana AI (+ Oxford, UBC collaborators)
- Status
- stable
- Spec
- https://sakana.ai/ai-scientist/
- Interop with
- Aider, OpenAI, Anthropic, Semantic Scholar
Frequently asked questions
Is the AI Scientist producing publishable research?
Outputs are uneven: some experiments are novel and clean, others are re-derivations or flawed. The project's contribution is demonstrating the end-to-end loop, not claiming paper-quality on every run.
How does it run experiments safely?
The pipeline runs inside sandboxed containers with resource and network limits. Sakana publishes guidance on preventing runaway loops or unintended side effects.
What changed in v2?
AI Scientist-v2 introduced stronger agent-based reviewer critique, broader tree-of-experiments search, and improved LaTeX and figure generation — widening the range of topics it can tackle.
Sources
- The AI Scientist — Sakana — accessed 2026-04-20
- AI Scientist GitHub — accessed 2026-04-20