Four necessary conditions, one empirical test that operationalises them, and three validation tiers. This page is the citable reference for the definition published in Cell Genomics.
The definition below is quoted from Corpas, Guio and Fatumo, Agentic genomics: From pipeline automation to autonomous validation, Cell Genomics 6, 101305 (2026), doi:10.1016/j.xgen.2026.101305. The article is open access under CC BY-NC-ND. Short quotations are reproduced here with attribution; for the full text, the tables and the figures, read the article.
Agentic genomics is the use of autonomous AI agents, powered by large language models (LLMs) and operating within domain-constrained skill libraries, to discover, plan, execute, and iteratively refine multi-step genomic analyses, where the agent exercises runtime decision-making over tool selection, parameterization, error handling, and output evaluation.
The article states that these four conditions are jointly necessary. A system meeting only a subset falls into one of the adjacent categories listed below rather than qualifying as agentic genomics.
The agent must make decisions during execution, not merely follow a static workflow specification.
The agent must operate within a structured library of validated operations (skills), not generate code ad hoc from a general-purpose model.
The agent must be capable of evaluating intermediate results and modifying its approach accordingly, including error diagnosis and self-repair.
The interface between the researcher and the analysis must be natural language, enabling researchers without programming expertise to direct complex analyses.
The conditions are meant to be checkable rather than argued about. The article gives one empirical test:
A system fails to qualify if, when presented with identical input under perturbed intermediate outputs (e.g., a tool returning an unexpected error or an anomalous result), it does not alter its execution strategy.
Two points about applying it honestly. The perturbation must be to an intermediate output during a running analysis, not to the input, because perturbing the input tests a different property. And the published criterion is binary: a system that halts safely without changing strategy fails it, exactly as a system that continues silently does. Those are different failures in practice, and separating them requires an extension to the published criterion rather than a reinterpretation of it. Any such extension should be declared as a change.
The article rules out four adjacent categories explicitly. Each is useful; none satisfies all four conditions.
| Category | Why it does not qualify |
|---|---|
| Workflow automation Nextflow, Snakemake, Galaxy | The workflow is specified in advance by a human developer. The manager executes a fixed graph; it does not decide at runtime which tools to use, nor adapt when intermediate results are unexpected. Agentic genomics builds on this infrastructure and is not a replacement for it. |
| AutoML and automated pipeline selection | Search over a predefined space of architectures or hyperparameters, within constraints set by a human engineer. It optimises within a search space rather than formulating the analysis strategy. |
| LLM-assisted scripting "vibe coding" | The model produces code, the human executes and evaluates it. No autonomous execution, no runtime adaptation, no domain constraint. |
| General-purpose AI copilots | Information retrieval systems that answer questions or summarise literature. They do not execute multi-step analyses against real data. |
The distinguishing feature, in the article's words, is that the agent acts: it executes tools, reads their output, decides what to do next, and delivers results. The researcher's role shifts from constructing the analysis to evaluating it.
The article proposes a tiered validation framework rather than a single pass or fail. The tiers are summarised here; the full criteria are in Table 2 of the article.
| Tier | In short |
|---|---|
| Research-grade | Reproducible and documented, with the analysis and its provenance recoverable. Suitable for exploratory research. |
| Benchmarked | Performance characterised against reference datasets by a party independent of the system's authors. |
| Clinical-grade | Validated to the standards required for clinical use, which are set by regulators and accredited laboratories. |
A draft protocol that operationalises the perturbation test, including fault classes, a reporting rubric, and the governance rules under which any assessment is published, is being developed in the open at ClawBio/ClawBench. It is a draft for comment and is not a standard. Comments are welcome as issues on that repository.
ClawBio's own skills are assessed first, and non-passing results are published. Under the tier definitions above, a system cannot claim the benchmarked tier on the strength of its own authors' benchmarking, and that constraint applies to ClawBio.
Cite the article, not this page.
@article{corpas2026agentic,
title = {Agentic genomics: From pipeline automation to autonomous validation},
author = {Corpas, Manuel and Guio, Heinner and Fatumo, Segun},
journal = {Cell Genomics},
volume = {6},
pages = {101305},
year = {2026},
doi = {10.1016/j.xgen.2026.101305}
}Corpas M, Guio H, Fatumo S. Agentic genomics: From pipeline automation to autonomous validation. Cell Genomics. 2026;6:101305. doi:10.1016/j.xgen.2026.101305
ClawBio is an open-source project founded by Manuel Corpas, who is a co-author of the article quoted on this page. A definition authored by the maintainer of a skill library, whose second condition concerns validated skill libraries, carries an obvious conflict. Three things are done about it rather than said about it: the terminology disclaimer above is the article's own and is reproduced in full; ClawBio is assessed against the definition before anyone else is, with failures published; and no ClawBio component is described as benchmarked on the basis of benchmarking performed by its authors. The article's full Declaration of Interests is published with it.