At I/O 2026 Google announced Gemini for Science, a collection of tools built explicitly for research workflows rather than for the generic knowledge worker. The interesting part is not a single product: it's the way DeepMind, Google Research, Google Cloud and Google Labs decided to ship together.
Three tools, one method
The first is Hypothesis Generation. Built on top of Co-Scientist (the multi-agent system DeepMind published in Nature), it simulates the scientific method: it works with the researcher to frame the challenge, then triggers an "idea tournament" in which agents generate, debate and score competing hypotheses. It does not answer a question. It opens a hypothesis space.
The second is Computational Discovery, built on AlphaEvolve and on ERA (Empirical Research Assistance). It's an agentic engine that generates thousands of code variants in parallel, scores them and surfaces the most promising. It targets the kind of problems where the gap between a mediocre and a strong answer comes from searching very widely.
The third is Literature Insights, built on NotebookLM. It crawls millions of scientific papers and produces tables, slides and navigable reports. It doesn't replace the reading, but it removes the days spent figuring out which papers actually matter.
Science Skills: the systemic piece
Then there's Science Skills, a bundle that integrates more than thirty life-science databases and tools: UniProt, AlphaFold Database, AlphaGenome API, InterPro and others. The Skills are available on GitHub and inside Google Antigravity, where a researcher can orchestrate structural bioinformatics or genomic analysis flows in minutes rather than hours. This is where the story gets interesting: Google is wiring its agentic developer platform into the actual workflows of working labs.
Why it matters
The generalist-AI narrative — "ask Gemini to write your thesis" — is a shortcut that doesn't speak to serious research. Here Google takes a different route: specialized agents, anchored to authoritative databases, built to accelerate the workflow of people who do science for a living. It's the first move toward a model in which AI doesn't answer, it collaborates with a method.