How the workflow is organized
The authors describe agents working on ideas, experiments, comparisons and report writing. A meta-controller manages parallel agents and checks each stage against a rubric, or set of criteria for judging its work. The repository’s instructions lay out six phases, from task drafting through investigation, final evaluation and paper writing, with optional review and revision phases.
Where humans stepped in
In Sakana AI’s account, researchers say they used an early version to investigate Error Diffusion, a biologically motivated learning rule. They set the question and constraints, chose further directions, and checked the implementations, results and references before submitting the resulting Diffusing Blame study. Sakana AI says the paper was accepted for an ALIFE 2026 oral presentation after external peer review.
Those checks mattered. Sakana AI says human inspection caught a high-scoring candidate that violated the intended learning rule, prompting the researchers to tighten their rubrics. The Agentic Scientist preprint’s abstract says the authors’ ablations suggest that reliable scientific autonomy depends on the coordination between agents as well as the underlying models.
Sakana AI reports evaluating 14 model-harness combinations across 70 campaigns on five research topics. It describes AI-reviewer scores as a proxy for paper quality, not independent evidence that the science is correct. The company says only the Error Diffusion paper underwent external peer review and that all its research tasks were in computational machine learning. For anyone trying the released code, the README warns that agents can run arbitrary commands and inherit supplied API keys; it recommends Docker, trusted tasks and sources, and reviewing run directories before sharing them.
