Biomni is a general-purpose biomedical AI agent developed by Kexin Huang and Jure Leskovec of Stanford University's Department of Computer Science, together with collaborators at Genentech, the Arc Institute, Princeton University, and UCSF, and published in Science on July 10, 2026. Unlike prior biomedical models specialized for a single task, Biomni autonomously breaks down a wide range of research problems across genetics, pharmacology, rare disease diagnosis, and more, composing tools without fixed workflows. On a database question-answering benchmark it reached 74.4%, essentially matching human experts at 74.7%, and on sequence reasoning (SeqQA) it scored 81.9%, exceeding humans at 78.8%. ASAP examines what this system changes and the limits of its numbers, based on the primary Science paper.

A Two-Layer System: Environment and Agent

The core of Biomni lies in separating an environment that holds the entire biomedical action space from an agent that reasons on top of it. The researchers first built an action discovery agent that mined essential tools, databases, and experimental protocols from tens of thousands of publications across 25 biomedical domains. The result is an environment called Biomni-E1, which integrates 150 specialized biomedical tools, 105 software packages, and 59 databases. The paper states that the discovered tools were rigorously verified by human experts.

Operating on top of it is the agent layer, Biomni-A1. Biomni-A1 combines large language model reasoning with retrieval-augmented planning and code-based execution: when a user query arrives, it first uses a retrieval system to find the most relevant tools, data, and software, then generates a detailed execution plan and composes the workflow itself. Not depending on predefined function-call orders or fixed templates, and instead assembling actions dynamically to fit the task, is where it diverges from earlier approaches.

What the Benchmarks Showed: Where It Caught Up to Experts

Biomni's benchmark performance varied sharply by task type, matching human experts on database questions while falling well short of them on the hardest reasoning problems. Across three multiple-choice benchmarks of general biomedical knowledge and reasoning, on the LAB-Bench database question-answering (DbQA) task Biomni scored 74.4%, essentially equal to the 74.7% of human experts and well ahead of the coding agent (ReAct+Code) at 40.8%. On SeqQA, which reasons over DNA and protein sequences, it reached 81.9%, surpassing the human level of 78.8%.

By contrast, on the hardest benchmark, the biomedical items of Humanity's Last Exam, Biomni reached only 17.3%. Even so, that was the highest relative score compared with the base LLM's 6.0%, the coding agent's 12.8%, and the literature agent's 12.2%. On eight realistic research tasks, including GWAS causal gene detection, perturbation screen design over roughly 20,000 genes, rare disease diagnosis, and drug repurposing, Biomni outperformed the base LLM without tool use (Claude Sonnet 3.7) by an average relative gain of 402.3%, the coding agent by 43.0%, and its own ablated variant Biomni-ReAct by 20.4%.

A Wet-Lab Benchmark: Competing With a Senior Researcher on Cloning

The most real-world test of Biomni's capability was a gene-cloning design task. The researchers, working with a group of gene-editing experts, designed an open-ended cloning benchmark and an expert user study, and posed the same tasks to four entities: a base LLM (Claude 3.7), Biomni, a human trainee at the level of a Stanford Biology master's with prior cloning experience, and a senior human expert, a Stanford Genetics postdoc with more than five years of cloning experience.

In blinded expert review, Biomni produced protocols and designs that matched the skilled human expert in accuracy and completeness, often providing comparable levels of detail and anticipating the same edge cases. The human trainee's submissions, by contrast, scored relatively lower in expert assessment. The paper summarizes that Biomni delivered accuracy comparable to human experts while completing the tasks in a fraction of the time.

How It Differs From Prior Biomedical AI: From Specialist to Generalist Collaborator

The significance of Biomni lies in a shift in biomedical AI's design philosophy from specialization to generality. Until now, the leading biomedical AI systems were single-purpose models that went deep on one well-defined task, such as protein structure prediction or gene expression prediction. Each is powerful on its specific problem, but none stitched together the full multi-step workflow a researcher actually performs, namely reading the literature, downloading data, choosing analysis tools, and interpreting results.

Biomni differs in approach because it is designed for a single agent to run through that fragmented workflow end to end. Mirroring how a human scientist alternates between retrieving knowledge and generating new insight, it repeatedly cycles between tool calls and reasoning to push a task forward. That is why the researchers frame it as a versatile collaborator, or co-scientist. It aims not to be a predictor that outputs one answer, but an execution agent that carries out the whole flow of research. This can be read as an attempt to elevate AI from an individual computational tool to an operator of the research pipeline.

How to Read the Numbers: What Impresses and What Warrants Caution

Biomni's scorecard holds both genuinely impressive results and figures that warrant caution, and separating the two requires reading each number in its proper context. Tying experts on database questions (74.4% versus 74.7%) and beating humans on sequence reasoning (81.9% versus 78.8%) shows that on well-defined tasks solvable through database lookups, the agent already reaches expert level. Competing with a senior postdoc on cloning design signals that this ability extends beyond pure knowledge Q&A into experimental procedure design.

Some figures, however, must be read in context to avoid misunderstanding. The 402.3% relative gain across eight tasks looks dramatic, but it also reflects that the comparison baseline, the base LLM without tools, had very low absolute performance. Relative improvement grows larger the lower the starting point, so the absolute level reached matters as much as the size of the gain. Scoring just 17.3% on Humanity's Last Exam honestly reveals that Biomni is not expert-level on every task, but varies widely by task type. It is strong on lookup-solvable problems, yet still has far to go on the hardest problems that demand deep reasoning.

Implications for Korea's Bio and Medical Research

The practical implication for Korean research is the potential to unlock data-analysis bottlenecks. In Korea's bio, pharma, and hospital research, tasks like GWAS analysis, drug repurposing, and rare disease diagnosis depend heavily on specialized personnel and tool access, and an open agent environment that integrates tools and databases, as Biomni does, has the potential to lower that barrier. Especially at smaller research institutions or clinical settings short on specialized analysts, the value of a collaborator that handles repetitive, labor-intensive workflows is far from small.

Yet any discussion of domestic adoption must also weigh the gates of validation and regulation. Because biomedical research outputs can lead to patient diagnosis and treatment, practical standards are needed for how far to trust an agent's outputs and where to require human expert verification. It is safer to accept Biomni for what it is: not a finished autonomous researcher, but a tool that shows its strength when it works alongside human experts.

Limits and Open Questions

Biomni's limits are acknowledged most clearly by the researchers themselves. The paper states that "Biomni has not achieved expert-level performance across all task categories," adding that no system yet captures the full scope of human biomedical expertise. The fact that the tools' accuracy was verified through human experts shows that the environment's reliability still rests on human oversight.

Three questions remain. First, how far can the ability that tied experts on lookup tasks extend to higher-order tasks like deep hypothesis generation and creative experimental design? Second, how will research settings control the hallucination and reproducibility problems that can arise as the agent composes tools on its own? Third, can the scores shown in benchmarks and controlled user studies hold up in the regulatory environment of actual clinical and drug development work? The researchers expect performance to keep improving as foundation models advance, the agent environment expands, and experts deploy Biomni in practice. ASAP will keep verifying the next advances in biomedical agents against primary papers.

---

Sources: Kexin Huang et al., "Autonomous biomedical research with an artificial intelligence agent," Science (July 10, 2026), DOI 10.1126/science.adz4351 · Biomni project paper PDF (Stanford)