Google wants agents to propose hypotheses and write experimental code instead of the scientist

Google Research at I/O 2026 was not just another AI demo set. The clearer message is that research work is moving from a chat assistant toward an agentic system that proposes hypotheses, writes experimental code, evaluates results and assists with literature. ERA and Co-Scientist are not added Gemini features, but an attempt to reorganize the researcher's work cycle.

ERA and Co-Scientist target the middle steps where a researcher's mechanical work accumulates

Google summarized several research tools around Gemini for Science. Empirical Research Assistance (ERA) is a system for writing and optimizing empirical research software, with examples from neuroscience to forecasting seasonal runoff across California river basins. Co-Scientist is a multi-agent system that generates, evaluates and refines hypotheses.

Google adds Computational Discovery, Hypothesis Generation, Literature Insights, Science Skills for agentic coding, and an experimental Paper Assistant Tool for peer review. The intent is clear: bring AI into the middle steps of research where the most mechanical work lives, not just as a chat partner for ideas.

Science has a problem exactly where the agent promises to help

A researcher has to read literature, write code, design experiments, run variants, inspect results and defend the conclusion. If an agent can safely iterate over hypotheses and code, it can change research speed in domains where the bottleneck is experimental setup and evaluation.

That matters more than the Gemini branding. But it only holds if the agent output is auditable and the researcher understands it well enough to see where the agent may have failed.

The agent generates hypotheses and elegant mistakes alike: without an audit they are indistinguishable

The biggest risk is false certainty. A scientific agent that generates hypotheses can also generate elegant mistakes. Citations, peer review assistance and scoring metrics help, but they do not replace the responsibility of the lab, the reviewer and the domain expert.

The second concern is access and vendor lock-in. If the best scientific workflow tools are tied to one vendor's cloud, models and platform, research infrastructure moves closer to a commercial stack. For academic institutions with open-access commitments, this is a serious question.

The signal will be a reproducible result outside the Google blog and a clear line for the Paper Assistant

Watch whether Google shows reproducible results outside its own blog and whether researchers get tools they can audit. For the Paper Assistant Tool, the key issue is whether conferences such as NeurIPS, ICML or STOC maintain a clear line between helping authors and invisibly influencing how work is judged.

For Gemini for Science, the test is whether these tools actually shorten the path from hypothesis to verified result, or merely produce more text and more variants that scientists have to clean up.

Lilith's verdict

Google is not just giving scientists a smarter chatbot here. It wants to build a lab where the agent writes the protocol and the human still has to watch for an elegantly formulated mistake sitting on the bench.