Google Research unveils Gemini for Science, ERA and Co‑Scientist at I/O 2026 to accelerate research

At Google I/O 2026, Google Research unveiled Gemini for Science alongside two experimental, agentic research systems — Empirical Research Assistance (ERA) and Co‑Scientist — designed to speed end‑to‑end scientific workflows. The announcements aim to reduce the time it takes to move from idea to tested model by automating hypothesis ideation, code generation and large‑scale computational experimentation. Empirical Research Assistance (ERA) is a research coding system that proposes concepts, writes and evaluates code, and iterates automatically to optimize empirical software. Google says ERA searches thousands of code variants using a tree‑search approach to find higher‑performing solutions.

Google also published results and code in a GitHub directory and highlighted case studies such as predicting hospital admissions for respiratory illnesses and forecasting seasonal runoff across California river basins. Co‑Scientist is a multi‑agent system built on Gemini that acts as a collaborative AI research partner. It uses a coalition of specialized agents that iteratively generate, evaluate and refine hypotheses, running multi‑agent workflows to surface and stress‑test candidate ideas. Foundational research on Co‑Scientist was also published in Nature, accompanied by researcher testimonials and prior validation papers that show applications across a range of domains.

Both ERA and Co‑Scientist are integrated into the broader Gemini for Science suite, which Google developed with teams across the company. One prototype called Computational Discovery pairs ERA with AlphaEvolve to form an agentic research engine that generates and scores thousands of code variations in parallel, enabling researchers to explore multiple hypotheses and modeling approaches that would otherwise take months to evaluate. Another prototype, Hypothesis Generation, uses Co‑Scientist to run ‘‘idea tournaments’’ where agents collaborate with scientists to define challenges, propose candidate hypotheses, debate alternatives and rank outcomes.

Google framed the releases against the practical difficulty that millions of papers are published each year, which makes literature synthesis and hypothesis ideation a major bottleneck for many fields. The company says these systems can accelerate workflows from automated idea generation to computational experimentation and cited application areas where ERA and Co‑Scientist have already been tested, including antimicrobial resistance, plant immunity, liver fibrosis, neuroscience and cosmology.