What we learn running lab science with agents: protocols, plate data, agent design, and how we work. Founder essays live at Building Elnora; company announcements at News.
Computational design searches an enormous set of possible molecules and returns the few worth making. An assay development AI assistant plans the experiment that settles them, because prediction is strong on chemistry and ordinary on the disease biology.
An AI model proposes ideas from a published record skewed towards success. An AI agent for drug discovery lab work earns its place on the validation experiment: designed well, run at the bench and kept, negatives included.
AI does speed up scientific research, by close to seven hours a week on scientists' own count, so the slow step moves to the bench experiment and to checking the output, and hypotheses wait in a growing queue.
An AI assistant for wet lab scientists belongs on the searching and the drafting, and the anomaly stays with the scientist, because odd results are where the field's biggest discoveries started.
Roughly 950 AI agents searched public DNA databases for 21 hours and flagged an enzyme system in viruses that infect bacteria. The searching is what an AI agent that runs bioinformatics pipelines from public data does, and Elnora does it for a lab's questions.
World models learn how an environment behaves and predict what follows an action. A knowledge base for lab protocols and results is the version a lab already owns, and Elnora reads it, ranks explanations from it, rehearses each run and files what comes back.
An AI lab protocol generator turns a described experiment into a plate map, a materials list with resolvable catalogue numbers and a set of deck-specific instructions, while a scientist keeps deciding what the experiment should test.
Protein structure prediction, image analysis and AI-written research code are already real. A lab that wants to automate lab data analysis with an AI agent needs one that reads the files its own instruments write.
A scientist's job is a bundle of tasks that are changing at different speeds, and the labs that adopt now, while still teaching the underlying skill, are the ones that keep both the speed and the judgement.
Vendors who want a product named in a protocol have to give Elnora the same evidence a scientist would check: datasheets, application notes and the runs behind them, including the ones that failed.
How do scientists use AI agents in the lab? Increasingly to run computational biology, quality control and design of experiments work outside their formal training, as long as they still supply the judgement that decides whether the answer is right.
An answer key only covers the questions that have one. How to evaluate an AI agent for scientific work when two good scientists would disagree, by scoring provenance, negatives, and the judge itself.
Every list of the best AI tools for biology research is written from demos. The method we run on Elnora: synthetic data from a model you control, a withheld answer key, and defects you never mention.
Scientists open the CSV their machine wrote and prepare it with colour codes and tabs. Handing that to an AI tool to analyze plate reader data passes on the answer you already expected.
The loop behind every document at Elnora, a shared knowledge base, an agent that drafts from it, a review pass and the result filed back, is how Elnora works as an AI agent that drafts lab reports from your data.
