DeepMind opens Co-Scientist, its multi-agent hypothesis system

DeepMind is opening Co-Scientist, the multi-agent system it built on Gemini to help researchers generate and test scientific hypotheses. The post describes a pipeline rather than a single model. A generation agent and a proximity agent propose and cluster initial hypotheses. A reflection agent runs peer review. A ranking agent then runs an "idea tournament" with Elo-style scoring borrowed from AlphaGo and AlphaStar. Evolution and meta-review agents refine the survivors, while a supervisor coordinates parallel exploration across several directions at once.

DeepMind says most of the compute goes to verification rather than generation. Agents cross-check claims against ChEMBL, UniProt, and web search, and the system can call specialised tools like AlphaFold during a debate. The post ties Co-Scientist to six published case studies, including a liver fibrosis result where a proposed drug repurposing blocked 91% of the scarring response in lab tests, plus ongoing work on ALS, cellular aging, and metabolic disease. Enterprise collaborators named in the post include Daiichi Sankyo, Bayer Crop Science, and US National Laboratories.

What is new this month is access. Hypothesis Generation, the experimental tool fronting Co-Scientist, is opening to outside researchers, with an enterprise version through Google Cloud. The framing is that the model has stopped being a hidden lab project and is now something a researcher can point at their own literature review.

Why it matters

If you run any research team that lives on literature synthesis, this is the first general-purpose hypothesis tool with named published results and named partner labs rather than another tech demo. Use it as an idea filter, not an oracle, and read the case studies before you trust its rankings on your own field.