How one scientist covers a team's worth of expertise

How do scientists use AI agents in the lab? Increasingly, for work that used to sit outside their own training. A molecular biologist can ask Elnora to run a variant-calling pipeline against a public sequencing accession, check a dose-response fit the way a quality-control specialist would, or have a proposed experimental design reviewed for a weak control before a plate is poured. A scientist directing this work stays a scientist: the computational biology, the quality control and the design of the experiment are the agent's execution, and the judgement that calls each one right or wrong stays theirs.

This is the part of the vibe coding conversation that gets skipped most often. Describing what you want in plain language produces something worth trusting when the person doing the describing already has taste, the ability to tell a sound result from a plausible-looking one. A scientist who has run the same three assays for a decade has exactly that kind of taste, just not for computational biology, quality control or design of experiments. An agent that can execute in those fields, while the scientist keeps deciding what counts as a good answer, is what lets one person cover ground a small team used to cover.

Specialists, without hiring them

Elnora divides its own work the way a well-staffed lab already divides people into roles: one subagent works the literature, another takes the numbers, and a third checks the experimental design for weak points and ranks them before a scientist ever reads the draft. That is the shape of a team a single investigator used to need to assemble. Directing all three from one seat still leaves the job that stays with the scientist alone: judging whether the flagged control actually matters for this particular assay, or whether a fit makes biological sense against everything else they know about that system.

The same pattern runs through the connections to public data. Single-cell quality control and variant-calling pipelines execute against accessions from GEO and SRA inside the same conversation that drafted the protocol, so a bench scientist with no bioinformatics background can request a named analysis and read the output next to the method it came from. Gene dependency scores pulled from DepMap narrow to a single cell line or tissue type, and target-disease evidence scored from Open Targets comes back with the drugs known to act on that target and the trial stage each reached, because deciding whether a score actually applies to the system in front of you is still a judgement call.

Design of experiments, checked as you write

The same holds for the ground a design-of-experiments specialist would normally cover. Buffer choices are checked against the salts they are paired with while a protocol is drafted, so an incompatible pairing is visible before anything is weighed out, and a hypothesis-generation skill proposes candidate explanations for a result and ranks each by the evidence behind it. The scientist still decides the experiment: picking which of the ranked hypotheses is worth the next attempt is exactly the decision a statistician gets hired to make on a bigger team.

What one scientist keeps, after doing the work once

A shortcut worked out this way turns into something the rest of the lab can use. Elnora lets a scientist author their own skills and subagents and wire them to outside tool servers, and it carries what they saved, the skills they built and the protocol templates they named, forward into whatever they open next. A pipeline one person set up for pulling variant calls, or a scoring pass someone wrote for a particular assay, gets a name the rest of the lab can call on, the way a shared protocol already works.

The wider discussion about vibe coding keeps landing on the same point from different directions: the people getting the most out of it already had strong judgement before they typed a word, and that gap widens as the tool gets more capable. An agent that can run the pipeline, fit the curve or flag the bad control extends what one scientist can take on alone. Reading what came back and knowing, from everything else they know about the system, whether it is actually right, stays the scientist's job the whole way through.