The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery

The short version

The problem

Coding agents can write programs and run experiments, but that alone does not make research reliable. Researchers need ideas tested fairly, results checked against the methods, and claims tied to evidence before they can trust or communicate findings.

The idea

The paper builds a workflow that gives different agents distinct jobs: propose a testable idea, implement and check it, share useful findings with parallel workers, and later run comparisons and analyses on the strongest result. Each step is checked against explicit requirements, and the final paper is assembled from saved experimental evidence. For scientific diagrams, a companion tool first makes an editable diagram and only then uses image generation to polish it, checking the polished version against the original.

What it got

The system supported a human-supervised paper accepted for a conference and produced autonomous research results across several computing tasks. In one key transfer test, its learned optimizer matched or beat a standard optimizer on 5 of 7 previously unseen control tasks; its figure tool scored 37.5 versus 25.0 for the best comparison system.