Co-Scientist: accelerating scientific discovery with a multi-agent AI system

Workflow & human oversight

Coordinated agents generate, critique and refine hypotheses using evidence within human-defined objectives. Humans perform experimental validation; the documented scope is not autonomous execution of an entire research cycle.

Workflow level is not an evidence score; higher is not better evidence.

Problem

Scientific discovery depends on generating hypotheses for complex problems that then undergo rigorous experimental validation. The authors ask whether a multi-agent system with structured reasoning and scalable test-time compute can produce genuinely novel, verifiable research hypotheses.

Approach

Co-Scientist uses a multi-agent architecture with an asynchronous task-execution framework for flexible compute scaling, plus a tournament evolution process in which hypotheses compete, are critiqued and are refined for self-improvement. Conditioned on research objectives and prior evidence, it formulates novel hypotheses for experimental verification.

Outcomes (as reported by the originating team)

The authors report that test-time compute scaling continued to improve hypothesis quality, and that in three biomedical applications the system helped identify new drug-repurposing candidates and synergistic combination therapies for acute myeloid leukaemia that were validated through in vitro experiments (Nature, 2026).

Limitations

Validated examples are three focused biomedical applications; general scientific utility beyond these is not established. Validation is in vitro. Hospital-partner involvement means the work is not reproducible outside similar collaborations without the same data access. Not independently reproduced.

Reproducibility

The publication describes the architecture, evaluation and validation. No public code repository was verified at source-check time (the repository located during checking returned 404).

Adoption requirements

Access to the underlying frontier model series (Gemini) and sufficient test-time compute; close collaboration with a wet-lab or clinical partner capable of validating generated hypotheses; domain experts to formulate the research objective precisely.