Sakana AI releases The Agentic Scientist: researchers set experimental rules, while agents explore hypotheses and write code
Sakana AI releases The Agentic Scientist: researchers set experimental rules, while agents explore hypotheses and write code
On October 15, 2026, Sakana AI published code and a preprint describing a system for computational research in which software measures the results. The researcher turns a question into a task with a predefined evaluation procedure; agents explore alternatives in parallel, and the system saves code, run logs, and measurements for the paper.
In The Agentic Scientist, the researcher first supplies an initial program, rules defining acceptable solutions, and an evaluation program that measures the results. A coordinator program allocates a limited compute budget among coding agents working in parallel. Each agent works in a separate directory: it proposes a hypothesis, modifies the code, runs a computational experiment, and saves the results to inform subsequent stages of the search.
The authors then compare the best candidate with the baseline and conduct ablations: they disable individual components of the method in turn to measure each component’s contribution. In the preprint, the authors describe 70 sets of search runs across five machine learning tasks, using nine language models with four programs that coordinate the agents’ actions. In these runs, the choice of language model had a greater effect on the results than the coordination program. When preparing the paper, the system links each claim to code, a run log, a measurement, or a comparison; an AI reviewer then evaluated the manuscript.
The authors describe the current design as a response to weaknesses in the original AI Scientist. An independent analysis of the 2024 version found that five of twelve proposed experiments failed because of code errors, and four of seven generated manuscripts contained incorrect or fabricated numbers. The new design preserves run logs, baseline comparisons, ablations, and a map linking claims to evidence so that a person can trace how each conclusion was reached.
The program that evaluates experiments also needs to be checked: in an internal Google test in which agents tackled research tasks, some tasks proved impossible because of defects in the evaluation program, so the scores conflated agent errors with measurement errors. In the Error Diffusion case, human review rejected a highly rated candidate because it violated the intended learning rule.
Error Diffusion trains a neural network by sending the error signal from its output directly to its hidden layers. In this study, people defined the question, the baseline comparison, and the evaluation rules; agents proposed hypotheses, modified the program, ran experiments, and analyzed failures. After human review, the authors submitted the work for external peer review, and the ALIFE 2026 conference program listed “Diffusing Blame” as an oral presentation on August 18.
In this workflow, the researcher formulates the question, approves the experimental conditions, and checks the results, references, and final claims. The agents explore alternatives, while run logs and comparisons allow a person to trace how each conclusion was reached.