Sakana AI Releases 'Agentic Scientist': Multi-Agent Framework Automates Scientific Discovery from Hypothesis to Paper

On October 15, 2026, Sakana AI released "The Agentic Scientist," the third generation of its "AI Scientist" system. This framework coordinates multiple coding agents, such as Codex and Claude Code, to automate the entire scientific workflow from hypothesis proposal and experimentation to analysis and paper writing.

Compared to previous versions, the new release introduces parallel exploration mechanisms and validation based on pre-defined evaluation criteria (rubrics). Different agents assume roles such as proposing hypotheses, executing code modifications, evaluating results, and drafting the paper. An audit agent specifically checks for numerical inconsistencies and claims that exceed the scope of experimental evidence.

The company reports that in specific tasks, such as optimizing information extraction in Doc-to-LoRA, accuracy improved from 52% to 98%. However, this result is limited to artificially constructed documents, and effectiveness in natural language scenarios remains unverified. The source code is available on GitHub under the Apache License 2.0, including sample projects runnable on CPU.

While the system can generate paper drafts, Sakana AI emphasizes that human researchers must still define the research direction, set evaluation standards, and take responsibility for final conclusions. Currently, the system is primarily applicable to machine learning research in computer science, with physical experiments or studies involving human subjects falling outside its validated scope.