Gemini-Powered ERA Agent Automates Scientific Discovery Code

Google Research published its Empirical Research Assistance (ERA) framework in Nature today, detailing an autonomous system that uses Gemini to write, execute, and iterate on scientific software. ERA functions as an agentic research engine, exploring thousands of methodological code variations in parallel to test hypotheses and analyze complex datasets.

Originally introduced as a preprint last year, the system is designed to automate the creation of “empirical software”—the custom analysis pipelines scientists build to evaluate experimental data. Google has released the codebase on GitHub (google-research/era), allowing developers to implement its core search strategies directly.

Flat UCB Tree Search Architecture

The ERA system relies on a specialized tree search algorithm called Flat UCB Tree Search (FUTS), a variant of the PUCT algorithm. Given a structured problem description, a target success metric, and an input dataset, ERA systematically traverses a search space of potential methodological approaches.

The agent queries scientific literature to propose mathematical concepts, implements those concepts as executable code, and evaluates the pipeline against the target metric. By running these iterations in parallel, ERA compresses a software optimization process that typically requires months of human effort into hours.

Benchmark Performance

The Nature publication outlines ERA’s performance across six multidisciplinary datasets. The system consistently met or exceeded the performance of existing expert-designed models.

Domain	Benchmark / Task	Key Result
Genomics	OpenProblems V2.0.0	Discovered 40 novel methods; top solution beat ComBat by 14% by integrating BBKNN.
Neuroscience	ZAPBench	Predicted the activity of over 70,000 neurons, surpassing the previous top video-based 3D volume model.
Public Health	CDC Hospital Admissions	Matched or outperformed expert forecasts for flu, COVID-19, and RSV admissions.

Additional performance gains were logged in geospatial analysis mapping atmospheric carbon dioxide, seasonal runoff forecasting in California, and satellite imagery evaluation.

Ecosystem Integration

Coinciding with the publication, Google integrated ERA into several enterprise and research platforms announced at Google I/O 2026. The new Computational Discovery tool combines ERA with AlphaEvolve to let researchers test modeling approaches at scale, currently available through a trusted tester program in Google Labs.

ERA also anchors the new Gemini for Science suite, which pairs the coding agent with Hypothesis Generation capabilities (powered by AI Co-Scientist) and Literature Insights via NotebookLM.

For enterprise bioinformaticians, Google Antigravity introduced over 30 life science databases, including UniProt and the AlphaFold Database. These resources are exposed as agent skills, allowing the system to execute complex genomic and proteomic analyses natively.

If you are building domain-specific reasoning engines, the FUTS reference implementation in the ERA repository provides a production-tested blueprint for evaluating AI agents against strict mathematical or empirical targets rather than relying on standard LLM-as-a-judge heuristics.