Run several drug programs with a few scientists and one record

A biotech can run multiple drug programs with a small team once an AI agent does the lab work for each one, and the programs gain most when they sit together in one lab with one shared record, so each program plans its next experiment from what the others have already measured. Take a team of three scientists carrying four early programs, each with its own target, its own cell-based potency assay and its own counter-screen for toxicity. Elnora, the AI agent for lab work, plans each program's experiments, writes and optimises its protocols, analyses its results and files them where the other three programs can read them, and the scientists choose which programs go forward and own the science behind each choice.

A drug program used to need a department and now needs a few people and an agent

Running one early drug program takes a long list of separate jobs: reading the literature on the target, designing the assays, writing the protocols, setting up the plates, fitting the dose-response curves and writing the report for the decision meeting. A company running several programs used to hire a team for each, or build shared departments to serve them, and the shared departments struggled because each program, as it matured, wanted staff who knew its own assays.

An AI agent takes most of that list. When the agent drafts the protocol, writes the run file for the instruments and analyses the plate that comes back, a program becomes a scientist who knows the biology working with an agent that does the lab work, and the agent gives each program its own attention while it works across the four.

Programs that share one record learn from each other's failed runs

Once a program can be run by a few people, splitting programs into separate teams or separate companies buys little, and it costs the data each program could have read from the others. Four programs in one lab often use the same cell lines, the same plate reader and the same counter-screen, so the run that failed in one program is often the run another program is about to set up.

Kept together, the programs pool more than equipment. The seeding density that worked for a cell line, the formulation that precipitated, the reference compound that drifted between batches and the reason a series was dropped sit in one place, and the fifth program starts from the four before it.

A team that owns a program stops it on the evidence

A small team that owns a program makes the go/no-go decision itself, and makes it quickly, because the program is theirs and the next one is waiting for the bench time. A program with a clear owner gets a clear decision, and the money and the plates move to the next program on the list.

The agent's part is putting the evidence for that decision in front of the owner the day it exists: each potency curve with its confidence interval, the counter-screen beside it, and what the other programs saw with similar compounds.

The bench sets the pace, so each experiment has to count

Software can predict that a molecule binds its target, and the potency assay, the toxicity counter-screen and the animal study still happen at a bench, in real time. Three scientists running four programs have a fixed amount of bench time, so the gain comes from designing each experiment well the first time and reading its result in full.

Elnora plans each program from what the others measured

Elnora is the AI agent for lab work, and for a team running several drug programs it plans each program's experiments, writes its protocols for the lab's own instruments, analyses its results and files them in a shared Knowledge Base that the next plan for any of the programs reads first.

Say the scientist leading the fourth program asks Elnora to plan a potency assay for a new series in a cell line two other programs already use, on at most two plates. Elnora searches the shared Knowledge Base with the whole question, and each passage it brings back names the file it came from: here the first program's runs in this cell line and a failed run from the second program where the signal collapsed at high cell density.

Elnora drafts the protocol to the group's named template, with the seeding density from the run that worked, and a critic agent reviews the design, ranks each problem it finds, such as a reference compound missing from the second plate, and proposes a concrete fix. Elnora then writes the files the connected liquid handler and plate reader need for both plates and rehearses the run in simulation, and the scientist approves both plates before anything is pipetted.

When the plates are read, Elnora fits the dose-response curves with confidence intervals in its sandbox, an isolated workspace where it runs code, and saves the figures under the name the scientist chose. The results land in the fourth program's folder in the shared Knowledge Base and the file tree updates across the organisation straight away, so the scientist on the first program sees the new data in that cell line the same day.

If the counter-screen shows toxicity rising with potency across the series, Elnora's hypothesis-generation skill lists the likely causes, from a shared off-target to a solubility problem, ordered by how much of the data supports each one, and the team decides whether the series goes on. A series stopped on that evidence stays in the Knowledge Base with its protocols and its reasons, ready for the next program that reaches for the same chemistry.

Role-based access sets who sees which program's files, a file a scientist attaches to a conversation stays out of the shared Knowledge Base until they file it, and the audit log records whether a person or the agent acted. The scientists choose the targets, approve each run and decide which programs carry on, and Elnora does the planning, the protocols and the analysis for each of the four.