Google DeepMind has introduced Co-Scientist, a multi-agent AI system designed to help researchers generate, test and refine scientific hypotheses. The company described the tool as a collaborative research partner built with Gemini, aimed at speeding discovery in life sciences and other fields.
The announcement came alongside a paper published in Nature and marks the latest step in DeepMind’s push to apply AI to scientific work. Google says the system will be made available to individual researchers through an experimental tool called Hypothesis Generation, with a rollout expected in the coming weeks. Researchers can register interest through a Google Labs site.
DeepMind said Co-Scientist is intended to address a familiar bottleneck in research: turning large amounts of literature and data into promising ideas worth testing in the lab. According to the company, the system is built around multiple specialized agents that work together rather than a single model producing a linear answer.
The process is divided into three main phases. First, generation agents propose possible research directions grounded in scientific sources. A proximity agent then organizes those ideas to broaden the search and prevent the system from focusing too narrowly. Next, reflection agents critique the proposals in a role similar to peer review, while ranking agents run comparisons and simulated debates to identify the most promising options. Finally, evolution and meta-review agents improve the strongest ideas and synthesize them into a proposal for a scientist to review.
Google said a supervisor agent coordinates the system and divides broad research goals into smaller tasks that can be run in parallel. The company argued that this approach differs from more traditional AI systems that move through problems step by step.
A major part of the system’s work is devoted to verification. DeepMind said Co-Scientist checks claims against published literature and data to improve factual accuracy and logical consistency. It also uses web search, databases such as ChEMBL and UniProt, and, in some collaborations, specialized tools including AlphaFold.
The company said the system can explore thousands of possible directions, then narrow them through what it calls a tournament of ideas. That approach, it said, takes inspiration from the strategy behind AlphaGo and AlphaStar, but applies it to scientific debate rather than games.
Google has already been testing Co-Scientist with research groups over the past year. The examples cited by the company focus largely on biology and medicine. In one case, researchers studying liver fibrosis said the tool helped surface drug-repurposing candidates, including one compound that blocked 91 percent of a scarring-related response in laboratory testing. Other collaborations have involved ALS, cellular aging, liver disease mechanisms, infectious diseases and aging research.
The company also said it has been previewing an enterprise version with organizations including Daiichi Sankyo, Bayer Crop Science and U.S. national laboratories involved in the Genesis Mission.
DeepMind emphasized that the system is meant to support, not replace, human scientists. Researchers quoted in the company’s announcement described it as a way to organize thinking, identify overlooked connections and narrow down the best questions for lab work. Google said Co-Scientist was developed with input from more than 100 institutions as part of efforts to build responsible and practical AI tools for science.