Google DeepMind Co-Scientist Enters Real-World Science

The important shift is from impressive demos to verifiable workflows—but this remains expensive, semi-autonomous research infrastructure rather than an unsupervised scientist.

• –Agent_H uses an eight-stage pipeline combining triage, decomposition, parallel candidate generation, tournament selection, clinical auditing, citation checks, and length optimization.
• –It required roughly 40–80 LLM calls per medical query, showing that inference-time scaling can improve capability without modifying model weights.
• –Physician evaluation found a meaningful safety improvement, but no significant quality advantage across eight other dimensions; automated judge scores should therefore be treated cautiously.
• –Real-world experiments included semi-automated CVD synthesis and lab-validated bacterial phenotype prediction, but human researchers still handled physical samples and oversight.
• –The system’s full source code is not public, limiting reproducibility and making cost, latency, and independent validation the next major hurdles.