Stanford open-source AI agent runs biomedical lab research

Clinical takeaway: An AI agent can now handle multi-step biomedical analyses at expert-level accuracy on defined tasks, pointing toward faster literature review, data analysis, and diagnostic support.

Biomedical data often piles up faster than anyone can analyze it. Much of it sits untouched because the work is slow, fragmented, and dependent on painstaking expert analysis. A new open-source agent from Stanford aims at opening up that bottleneck. It surveys the scientific literature, writes and runs its own code, and executes multi-step research workflows from a plain-language request. On benchmarked tasks it reached expert-level accuracy in minutes rather than hours.

The system, called Biomni, was tested across more than 400 research tasks. It matched senior-scientist performance on rare-disease diagnosis, identifying disease-causing genes, and drug repurposing, while cutting analyses that take hours down to minutes. The agent works by retrieving the relevant tools and databases for a given question, writing code to run them, and revising its plan as results come in.

In one hands-on test, Biomni designed a full molecular cloning protocol, including guide RNAs, primers, and a plasmid map, for a CRISPR editing task. Researchers ran it in the lab exactly as written; the cloning succeeded on the first attempt. The authors note this was a single demonstration rather than a systematic benchmark.

Beyond benchmarks, the team ran five case studies. In one, a user uploaded more than 450 files of continuous glucose, diet, and activity data and asked Biomni to find plausible hypotheses; it cleaned and unified the data and surfaced patterns in about 40 minutes. Other cases had the agent optimize protein stability and connect to wet-lab instruments. The authors also describe early efforts to extend Biomni to tasks like protein design and driving lab automation, though they call these preliminary.

To build Biomni, an "action discovery" agent mined about 2,500 recent papers across 25 biomedical domains to assemble a shared workspace of 150 specialized tools, more than 100 software packages, and 59 databases. Performance was measured against standardized benchmarks and, on several tasks, against human experts.

The benchmarks cover only part of biomedical research, and the authors flag that key domains remain untested. Physical experiments are still largely out of reach, with only early links to lab robotics. And because the agent can produce plausible but wrong outputs, its makers stress that a scientist has to verify the work, which is where the near-term value sits: faster analysis with human judgment still in charge.

"This is not about machines taking over science, but more about machines becoming a powerful new partner to augment human researchers. With Biomni, scientists have a fast and tireless collaborator that empowers them to focus on the important work of science," said Kexin Huang, a former doctoral student in Jure Leskovec's lab at Stanford University who co-founded a startup to bring Biomni to market.

Source: Huang K, et al. (2026 Jul 9) Science. Autonomous biomedical research with an artificial intelligence agent