No. 016

Two weeks ago the story was product launches: Claude Science, Google's Paper Assistant Tool, a DIY Claude skill. This week it is peer-reviewed publications and editorial recognition. Science published Biomni. Nature published Co-Scientist. Nature News published a survey treating the whole cluster as a settled category. The gap between "vendor demo" and "institutional endorsement" closed fast. Meanwhile the open-source alternative shipped and the first AI-quality-ranking system scored the entire bioRxiv corpus, which means the verification layer is already behind.

AI Co-Scientists

• Meet Biomni, an AI-powered biomedical co-scientist

  Stanford Report, July 9 2026

  A general-purpose biomedical AI agent published in Science after a year of peer review, reporting adoption at 10,000+ labs, making it the most widely deployed AI co-scientist in biomedicine and the first in this category to clear the peer-reviewed-plus-real-adoption bar simultaneously.

• Co-Scientist: A multi-agent AI partner to accelerate research

  Google DeepMind, July 9 2026

  DeepMind's Gemini-based multi-agent system for hypothesis generation, debate, and ranking published in the same issue of Nature, with a drug-repurposing demonstration that blocked 91% of a scarring-linked response in liver fibrosis lab tests.

• Built with Claude: Life Sciences

  Cerebral Valley, July 7 2026

  Anthropic's first post-launch hackathon for Claude Science selected 500 from 6,000 applicants, with Gladstone Institutes contributing live research datasets (T-cell sequencing, DNA regulatory-activity predictions, protein interaction networks), an institutional-endorsement signal distinct from sponsorship.

Tools and Infrastructure

• OpenScience: the open-source AI workbench for scientific research

  Synthetic Sciences, July 3 2026

  Apache-2.0, model-agnostic, local-first research workbench that runs the full loop (literature, hypothesis, code, experiments, write-up) and gained 880 GitHub stars in four days, positioning itself explicitly as the open alternative to Claude Science.

• PaperBanana: open-source implementation of the academic-illustration agent

  llmsresearch, July 2026

  MIT-licensed reimplementation of the five-agent Google Cloud AI Research and Peking University framework (Zhu et al., arXiv 2601.23265, January 2026, 72.7% win rate on methodology-diagram generation), extended to slide generation and shipping CLI, Python API, Gradio UI, and MCP server integration for Claude Code and Cursor. Figures are no longer the researcher's least-automatable deliverable.

• Generative and Agentic AI for Biology Workshop

  ICML 2026, July 10 2026

  ICML's dedicated workshop on autonomous biological AI systems ran in Seoul this week, covering agent-based hypothesis generation, closed-loop wet-lab integration, and benchmarks for autonomous scientific systems, as the conference reported "agentic AI" in at least 60 of 247 workshop proposals.

Framing and Accountability

• Which 'AI scientist' suits your lab? A guide for the perplexed

  Nature News (Ewen Callaway), July 10 2026

  Nature's editorial apparatus now treats AI-scientist tools as a surveyable category rather than isolated launches, with the sharpest line being: "faster discovery without rigorous validation simply allows us to produce errors more efficiently."

• Can AI tools spot great science before reviewers do?

  The Scientist, July 2026

  QED Science scored all 57,455 bioRxiv preprints from a 12-month window on originality and validity (AUC 0.867), but its published "top 1%" still concentrates at well-funded institutions, raising the question of whether algorithmic quality metrics measure science or resources.

• Ethics journal retracts paper by high school student for AI, peer review manipulation

  Retraction Watch, July 6 2026

  The Journal of Medical Ethics retracted a paper whose sole author used generative AI to "identify and understand referenced sources" without verifying them, producing fabricated citations, while hallucinated-reference rates across all journals jumped from 1-in-2,828 papers in 2023 to 1-in-277 in early 2026.