Co-Scientist (Google)
Gemini-based multi-agent reasoning engine focused on hypothesis generation, using a generation/reflection/ranking/evolution ensemble with Elo-style tournaments. Validated in three peer-reviewed biomedical case studies.
| Affiliation | Google Cloud AI Research / Google DeepMind / Google Research, with collaborators at Stanford, Imperial College London, and Houston Methodist (Google Research) |
| First introduced | 2025-02 (initial blog announcement); peer-reviewed paper accepted 2026-05 |
| Lifecycle stages | Hypothesis, Experiment design |
| Autonomy level | Semi-autonomous (scientist-in-the-loop collaborative paradigm) |
| Domain focus | General; validated in biomedicine (drug repurposing, target discovery, antimicrobial resistance) |
| Availability | Closed / API only — full source not public; experimental access program announced; Gemini foundation model accessible via Google APIs |
Approach
Multi-agent system built on Gemini, comprising specialized agents (Generation, Reflection, Ranking, Evolution, Proximity, Meta-review) coordinated via an asynchronous task framework. Uses self-play scientific debate for hypothesis generation, an Elo-style tournament to compare hypotheses, and an evolution loop to refine them by scaling test-time compute.
Validation
Three biomedical case studies:
- Drug repurposing for acute myeloid leukemia (AML) with in vitro confirmation.
- Novel epigenetic targets for liver fibrosis confirmed in human hepatic organoids.
- Recapitulation of a then-unpublished bacterial gene-transfer mechanism in antimicrobial resistance, discovered independently by collaborators at Imperial College.
Notable results
Identified novel single-agent and combination repurposing candidates for AML showing selective cytotoxicity at clinically relevant concentrations. Ranked epigenetic targets with anti-fibrotic activity in organoids. Independently recapitulated an unpublished AMR gene-transfer mechanism.
Primary paper
Gottweis et al., “Accelerating scientific discovery with Co-Scientist,” Nature 2026.
Other references
- Nature news on AI agents in science (AI Index 2026)
- arXiv preprint, “Towards an AI co-scientist,” 2502.18864
Code
Not released. Pseudocode and system prompts provided in supplementary notes.