ERA Published in Nature: AI Science Assistant Begins Automatically Writing and Optimizing Expert-Level Empirical Research Code
Google Research published the ERA Nature paper, code, and experiments in May 2026, and used it in a Computational Discovery prototype to explore and optimize scientific computing solutions via tree search.
AI for Science has moved from literature Q&A and hypothesis generation to executing empirical code. ERA can search methods, write code, run experiments, and compare results, allowing the scientific agent to be tested for the first time through both formal publication and runnable code.
ERA searches a large number of candidate solutions based on the problem and success metrics, combines existing methods, and iteratively optimizes programs. The Nature paper reports its expert-level performance on multiple types of empirical tasks; the official code is also publicly released, and the results are integrated into the controlled experiment product of Gemini for Science.
Scientific software, data analysis, and high-value R&D processes may form a new agent platform layer, but result credibility will depend on experiment reproducibility, data governance, domain expert review, and computational cost.
R&D organizations should start piloting scientific agents with verifiable small-scale computational experiments, using replication success rate and research cycle shortening as metrics, rather than evaluating by the amount of generated code.
Independent teams need to replicate experiments and observe real discovery rates, false hypotheses, code security, data leakage, and expert time savings in controlled trials.