Sakana AI releases The Agentic Scientist, an open-source coding-agent research framework

Sakana AI publishes The Agentic Scientist, a framework that runs computational research through coding agents.

Elvis Saravia says the system splits a project into stages such as ideation, planning, sandboxed experiments and report writing, with each stage run by a coding agent that checks its own work, and a meta-controller that runs many agents in parallel and checks every stage against a rubric.

Sakana AI says the framework wraps off-the-shelf coding agents such as Codex and Claude Code, that parallel workers test hypotheses against rubrics defined in advance, and that the company is releasing both systems as open source. Sakana AI and hardmaru also say that in autonomous runs the system raised Doc-to-LoRA retrieval accuracy on multi-needle tasks from 52% to 98%, and that a paper it helped produce on Error Diffusion was accepted as an oral presentation at ALIFE 2026. These are the sources' claims; the reports give no independent evaluation of them.