Let the agent that proposes an experiment also run it

Where AI helps in the wet lab now includes the bench itself: an AI agent can propose an experiment, write the run for the lab's connected instruments, read what those instruments report and propose the next step from that result. Take a molecular biology group building a reporter construct from a vector backbone and two inserts by Gibson assembly, where the design, the reaction set-up, the sequence check and the decision about what to do next are usually separate jobs done on separate days.

Elnora is the AI agent for lab work, and its job covers planning a lab's experiments, writing and optimising its protocols, analysing its results and remembering its decisions. For this group it designs the assembly, writes the run that sets up the reactions on the liquid handler, checks each clone against its sequencing trace and plans the next step from what came back, while the scientist chooses the construct and confirms the deck before the instrument moves.

A proposed experiment waits on the hand-offs between the plan and the bench

AI in science has mostly done the computational part: analysing data, making predictions and searching the literature. A model can suggest a construct or a molecule in minutes, and testing it takes a protocol, a program for the liquid handler, reagents placed on a deck, an instrument export and an analysis, usually done by different people at different times.

Each hand-off is a point where the experiment that runs can drift from the experiment that was proposed, for example when a primer is reordered with a shorter overlap and nobody updates the design file. The result then reaches the next plan once somebody carries it back, and the reasoning behind the original proposal may sit in a chat window that is long closed.

An agent that reads its own results chooses the next step

Conventional lab automation runs a protocol that somebody programmed, in the same steps each time, which suits a routine assay. An agent that writes the protocol and the instrument run from a goal, and then reads what the instruments reported, can change the next run because of the result: redesign a primer, add a control, drop a condition that gave no signal.

That works when the plan, the run and the result sit in one record. The next proposal is then made from a measured outcome, the runs that stopped early included, and a scientist can trace each choice in the plan to the result that prompted it.

The scientist chooses the question and confirms each run

The scientist decides what the construct is for and which result would settle the question, chooses which instruments and assays the agent works with, and confirms the physical deck before the hardware starts, because deck positions differ from one instrument to the next. Elnora does the lab work between those decisions: the design, the protocol, the instrument run, the analysis and the proposal for what comes next.

Elnora takes a three-part assembly from design to checked clone

The scientist asks Elnora to assemble the reporter construct and to plan the work so it ends with sequence-verified clones. Elnora reads the group's earlier cloning runs in the Knowledge Base, the lab's shared store of protocols and results, and designs primers with an overlap at each junction. It simulates the PCRs and the assembly and predicts the band sizes the scientist should see on the gel for each fragment.

A critic agent, a second agent that reviews the design, ranks each problem it finds and proposes a fix. Here it flags that the plan has no vector-only assembly, the reaction that shows how many colonies carry the backbone without the inserts, and its fix adds one to the plate.

Elnora then writes the run file the liquid handler executes to set up the PCR reactions and the assembly reactions, written against the deck that instrument has, and rehearses it in simulation, a software pass through each step before anything moves. The run stops to show the layout it expects, and the scientist confirms the bench matches. The run log comes back with the steps, their order and their durations.

When the Sanger traces for eight colonies return, Elnora checks each construct against its trace. Six match the design, and two carry the same deletion at the junction between the second insert and the backbone, so Elnora reports where the mismatch sits and proposes a longer overlap for that junction. It plans the next experiment too: an expression test of the verified clones on the plate reader, with the plate map and the reader's run file written, and the readings matched back to their wells when the plate is read.

The design, the run files, the logs, the traces and the decision land in the project's folder in the shared Knowledge Base, so the group's next assembly starts from the junction that failed and the overlap that fixed it.