Stanford's free Biomni turns plain-English questions into full lab workflows
The AI agent lands in Science with a 35 minute benchmark and 10,000 users, but the superlatives, the headline speedup, and a startup spinoff all need a closer read.
The AI agent lands in Science with a 35 minute benchmark and 10,000 users, but the superlatives, the headline speedup, and a startup spinoff all need a closer read.
A peer-reviewed biomedical AI agent is now in the hands of working scientists, and the paper describing it landed in Science this week. Stanford's Biomni takes a biologist's plain-English question and turns it into a full research workflow: searching databases, writing analysis code, finding disease-causing genes, and creating step-by-step lab instructions that human scientists have actually followed in real experiments.
The system is the work of a Stanford-led team whose two main architects were PhD students at the time, Kexin Huang in computer science and Yuanhao Qu in cancer biology, with computer science professor Jure Leskovec supervising. Huang and Qu built the system as an autonomous co-scientist: rather than answering one question at a time, Biomni plans a multi-step workflow, executes the steps through code, and pulls from biomedical databases and analysis tools, the way a junior lab member might.
That mechanism is what separates Biomni from the biomedical AI tools that came before it. Earlier systems tended to specialize: CRISPR/Coscientist for gene-editing design, ChemCrow for chemistry, BioGPT and its descendants for biomedical text. The team's framing of "first general-purpose biomedical AI agent" depends on where you draw the line between a specialist and a generalist, and the announcement post frames the contribution in those exact terms.
The benchmarks are broad but unevenly reported. The team tested Biomni on more than 400 real-world research tasks spanning wearable data analysis, genetic reanalysis, cloning protocol design, and lab instruction generation, comparing the system against existing AI tools and human experts. Most coverage has latched onto a single illustrative case: a complex biological data analysis that a human expert finished in roughly three weeks and that Biomni completed in 35 minutes at expert-level accuracy. That ratio is real and described by the authors as characteristic of this class of multi-step task. It is still one case inside a 400-task study, so the headline number should be read as the system's ceiling rather than its average.
The adoption story is bigger and softer. Leskovec told reporters this week that more than 10,000 scientists worldwide are already using Biomni for everyday research tasks. That figure comes from the supervising professor, not from an independent audit of the open platform's registration logs. It also sits alongside a code release rather than a peer-reviewed user study, so "10,000 registered users" and "10,000 scientists who trust the tool with their benchwork" are not the same claim.
Huang, one of Biomni's two main developers, has since co-founded Phylo, a San Francisco-based startup whose stated mission is to make the AI system available to more researchers. That is a familiar academic-to-commercial arc, but it sits inside a research project that is also being given away as open source from the same lab, which puts the conflict-of-interest question on the table from day one.
What still has to be tested is the part that does not show up in a benchmark: whether a workflow the agent designed in silicon actually holds up in a wet lab, whether the literature summaries it generates are reliable enough to act on, and whether hallucinations in molecular biology carry the same cost as hallucinations in less consequential domains. The open-source release is designed to invite exactly those tests, and the Science paper lays out the evaluation methodology that independent groups can rerun.
For now, the relevant milestone is that a general-purpose biomedical AI agent is no longer a thought experiment. It is a Science paper, an open-source release, and a claimed 10,000-user base, and the next round of evidence will come from labs that did not build it.