Executive Summary

A laboratory now exists that takes one request typed in plain words and carries a DNA experiment from design through execution by itself. AutoDNA, which researchers at Shanghai Jiao Tong University published in Nature on November 4, 2026, has six AI agents work the procedure out in conversation, write the code that drives the machines, and send it to more than 20 instruments.

The longest demonstration was a round trip that put digital information into DNA and read it back out. It came to 11,382 hardware steps over about 203.5 hours, a little more than eight days, and the original sentence came back intact. The phrase no human intervention carries a qualifier, however. People topped up reagents and consumables and cleared waste across those same eight days, and the companion article Nature published with it records that six of the fourteen hardware errors logged in the AutoDNA work needed a person's hands.

Sections 1 through 4 follow what the paper, the commentary beside it and the project documentation the team published will bear out. Section 5, which asks what a record made this way is worth to people who handle data, is this article's reading.

Key Figures

Sources: the Nature paper An AI-native experimental laboratory for autonomous biomolecular operations and, in the same issue, AI agents enter the 'wet lab'

The first two of the four cells are what AutoDNA managed, and the last two say under what conditions it managed them. Read without the second pair, the achievement looks larger than it was.

11,382

Hardware steps run without a person

The count for one DNA data storage round trip, which took about 203.5 hours, a little more than eight days

97.8%

Stepwise average yield in enzymatic DNA synthesis

On par with the results human experts have published after optimizing them

6

Hardware errors that needed a person

Of the fourteen logged, the system recovered from eight and six were left over

1

Clinical sample used for cancer mutation detection

It found KRAS G13D in a single piece of colorectal cancer tissue whose sequence it had not been given

1

Eight Days from a Single Line of Plain English

AutoDNA takes one line of natural language as its input. Everything after that follows in order: searching the literature, picking the reagents to use, laying out the experimental procedure, writing the code that drives the instruments, running them, and reading what comes out. By design there is no place where a person signs off on the procedure or touches up the code.

The longest of the four demonstrations is the DNA data storage round trip. Digital information is turned into a base sequence, which is the encoding step; DNA is synthesized to match that sequence; the material is stored; it is sequenced and read again; and the original data is recovered, which is the decoding step. AutoDNA tied those five stages together on its own, synthesizing and sequencing 78 DNA strands, and pulled back out the sentence that had gone in. The hardware operations came to 11,382 steps and the run took about 203.5 hours.

The DNA Data Storage Round Trip — Five Stages The same sentence goes in, passes through five stages, and comes back out the same Encoding data → base sequence Synthesis 78 DNA strands Storage kept as DNA Sequencing read again Decoding base sequence → data Same sentence — recovered intact 11,382 steps · about 203.5 hours · a little more than eight days
▲ Pebblous original diagram. The five stages and the figures are drawn from the text of the Nature paper.

The number speaks less to scale than to duration. Over eight days a robot arm carries samples between machines while heating, cooling, centrifugation and fluorescence measurement take their turns. One step going wrong partway through brings down the thousands that follow it. The paper puts the platform's positioning accuracy at the micrometer level and its liquid handling at the microliter level.

Here is where this parts company with the automated laboratories built until now. Earlier platforms added artificial intelligence on top of a workflow a person had already laid out. A model took on one task, data analysis or the choice of conditions, while the instrument lineup, the definition of the procedure and the optimization rules stayed in human hands. AutoDNA calls itself AI-native, meaning the laboratory was designed around the model from the beginning. Haibo Chen, a computer scientist at Shanghai Jiao Tong University and a co-author of the paper, described the difference as building a laboratory on the premise of artificial intelligence rather than adding artificial intelligence to a laboratory that already stood.

What makes the design worth attention comes out of a mismatch in speed. James Zou, the Stanford professor who wrote that commentary, noted that artificial intelligence can design candidate molecules with biological activity in minutes, while confirming them at the bench still takes weeks or months and no small amount of labor. Design got faster and the bench did not keep up. He sees AutoDNA as having begun to narrow the gap between proposing an experiment and actually carrying it out.

2

Six Agents in Conversation, and Code That Runs in a Virtual Lab First

The brains of AutoDNA are six agents with the roles divided among them. They fall into two groups, the side that settles what to run and how, and the side that turns that into movement of actual equipment. The experiment planning agent sits at the center and directs the whole.

2.1The Four That Settle What to Run

The first four never touch the bench. Their share of the work runs from searching the literature and counting what reagents remain, to laying out a procedure that stays inside that stock, to working out the next move when things do not go as intended.

  • Experiment Planner (EPA) — gathers the judgments the other agents return, lays out the experimental procedure, and directs the optimization as a whole.
  • Hypothesis Proposer (HPA) — when a result falls short of expectation, points to the cause and forms a hypothesis for improvement.
  • Literature Researcher (LRA) — finds and reads papers, and brings back the grounds used in designing the procedure and choosing conditions.
  • Reagent Manager (RMA) — counts the reagents on hand and keeps the procedure from going past that limit.

2.2The Two That Move the Instruments

The other two take the procedure the first group has agreed on and put it into a form the instruments understand. Two translations happen here: a plan written in natural language becomes Python code, and that code becomes machine motion.

  • Program Developer (PDA) — recasts instrument functions as atomic services, writes runnable Python code, and verifies it itself.
  • Hardware Executor and Validator (HEVA) — loads the code onto the instruments and runs it, confirms the results, and tracks in real time whether each machine is free or occupied.

The six make a closed loop. Design, run, look at the result, design again. When a result comes out poorly the hypothesis side offers an improvement and the loop turns one more time. The place where a person steps in lies outside that loop.

The closed loop of AutoDNA, with people outside it Six agents confer in natural language and repeat design, experiment and optimization The four that settle what to run Experiment Planning EPA · directs Hypothesis Proposal HPA Literature Search LRA Reagent Management RMA The two that move the instruments Program Development PDA · atomic services and Python code Hardware Execution and Validation HEVA · tracks machine status in real time procedure result Outside the loop, the human share Task selection and platform setup, reagent refills, waste disposal, and errors it could not recover from
▲ Pebblous original diagram. The agent names, their roles and the division into two groups are taken from the AutoDNA project page the researchers published. The diagram draws the agent structure only and leaves out the scheduler and the hardware control layer.

2.3Reading the Manuals and Breaking Machines into Atomic Services

For a model to handle an instrument it has to know what that instrument can do. Control methods and interfaces differ from one manufacturer to the next, so a system in which agents deal with low-level control machine by machine grows complicated quickly. AutoDNA puts one layer in between. It reads the specification sheets and the user manuals, divides each function into units that cannot be split any further, and attaches a written description and a set of parameter constraints to every unit. A single pipette becomes two services, liquid transfer and mixing. The agents combine these services into complex procedures.

Code that has been written does not go straight to the instruments. It runs once in a virtual laboratory first, which tracks liquid volumes and checks that vessels sit where they should and that instrument settings match. The example Zou raised in his piece shows what that stage is for. In one experiment the generated code tried to read a fluorescence value before the sample vessel had been placed in the machine, and the virtual laboratory caught the error first. It was filtered out before an experiment that costs real reagents and real time ever began.

When several users submit experiments at the same time, the system watches machine status and adjusts the order. If one instrument is occupied it reroutes to another with the same function and rewrites the control code to match, so that an occupied heater gets covered by a thermal cycler. The paper puts the execution time of concurrently running tasks at about a third of what it was under this arrangement. The engine is a multi-agent setup running on Google's Gemini API, and the code is public on GitHub under the Apache 2.0 license.

3

What the Four Test Cases Show

The team put AutoDNA through four tasks of differing character. They differ in kind rather than in difficulty. One asks for a fixed procedure to be followed exactly, one asks the system to feel its way toward conditions that have no settled answer, one asks it to carry a long process through to the end, and the last asks it to take an unknown sample and reach a conclusion.

Task What AutoDNA did Result
Nucleic acid detection Looks for African swine fever virus with a commercial fluorescent isothermal detection kit. The flow has three stages, reagent mixing, incubation at 39 degrees Celsius and fluorescence measurement, and once the reaction starts it repeats five minutes of incubation and a reading five times over Statistically indistinguishable from what a person produced with the same procedure
Enzymatic DNA synthesis Varies the buffer, the surfactant, the reaction time and the reagent concentration by itself, fitting yield and elapsed time together. No person sets the search range in advance Stepwise average yield of 97.8 percent, matching what human experts have published
DNA data storage Joins encoding, synthesis, storage, sequencing and decoding into one run. Synthesizes and sequences 78 DNA strands The original sentence recovered intact, in 11,382 steps over about 203.5 hours
Cancer mutation detection Takes a colorectal cancer tissue sample whose sequence it has not been given and works through five stages on its own, sequencing, locating the mutation site, primer design, primer synthesis and verification Found the KRAS G13D mutation and confirmed it again with primers it had synthesized itself

The second of the four stands apart from the rest. The other three have a clear goal and an order to follow. Optimizing enzymatic DNA synthesis was open from the start as to which variables even needed touching, and AutoDNA kept changing the direction of its search as it watched the results. The point of this demonstration is that it arrived somewhere close to the conditions people spent years refining before publishing them.

Put the four side by side and one thing is common to them. All of them work with nucleic acids, and their procedures are comparatively standardized. So the result sits closer to 'experiments of this kind can be automated all the way through' than to 'any experiment can be automated'.

4

How Far 'Autonomous' Reaches

'Autonomous' points at a narrower range than it sounds. Zou wrote plainly that choosing the experimental task and putting the physical platform together remain human work. Before an experiment starts a person prepares the reagents and the labware, and over a long run a person refills consumables and clears waste. The loop turns only when both are present, the person who submits the request and the person who built the laboratory.

Numbers on reliability were published alongside the paper. The researchers logged fourteen hardware errors over the course of the AutoDNA work, and the system recovered from eight of them by itself. The remaining six needed a person. Read as an automatic recovery rate it is 57 percent; read as the number of times a human was called in, it is six. Both describe the same fact, and for anyone reviewing a design that assumes unattended operation, the second is the figure that matters more.

The paper names limits of its own as well. Large language models carry randomness, which can affect the reproducibility of an experimental procedure. There is no guarantee that the same request yields the same procedure a second time, and in experimental science reproducibility is not a side condition. The paper also notes that the experiments suited to full automation today are mostly comparatively standardized assays, and that complex work involving animal models falls outside the range.

Zou pointed to the interface as the thing standing in the way of wider use. Programming interfaces differ from one manufacturer to the next, which makes an automation solution built in one laboratory hard to carry over to another. His warning is that unless an industry-standard interface comes to stand between the agents and the instruments, each autonomous laboratory ends as a custom build. The team named stronger instruments, mobile robots that carry reagents themselves, and a fuller literature database with retrieval-augmented generation as the tasks that come next.

5

Why Pebblous Is Watching This Paper

A laboratory running unattended means experimental data comes into being without passing through human hands. Those 11,382 steps are both work a machine did and a record a machine wrote. That record becomes training data for the next model, the grounds on which another team picks up the procedure, and the only trace left for anyone who doubts the result to follow back.

The most valuable part of what people used to write in a lab notebook was never the numbers but the reason for doing it this way. That place is not empty in AutoDNA. The six work things out in conversation, so capturing what they say to one another leaves something denser, in form at least, than a human notebook. Why the buffer was changed and which paper supplied the grounds both stay behind as sentences.

The team did not leave that record where it fell either. The data availability statement in the paper puts the sequencing data and the outputs the agents exchanged on Zenodo. Not only the results of the experiments but the conversation that led to them was brought inside the scope of release.

The trouble is that the limit the paper sets for itself cuts into the value of that very record. If randomness in the model can shake the reproducibility of a procedure, then the reason for choosing this condition as written in the log is either the causation that actually operated at the time or a plausible account composed afterward. The same suspicion attached to notebooks people wrote, but a person could be asked, and the author of a lab notebook answered for their own judgment. A generated account offers neither.

A small instance of this sits inside the news itself. For the same DNA data storage experiment on the same system, the team's project page reads 11,386 steps and 162.9 hours, while the published Nature version reads 11,382 steps and about 203.5 hours. The project page still points at the preprint, which looks like where the gap comes from. Neither record is wrong, but unless the version each belongs to is written down beside it, the two numbers look as though they contradict each other. A record written by a machine does not converge into one set of its own accord.

Anyone taking in automatically generated experimental data has four questions worth asking now. Asking before the records pile up costs a different amount of effort than asking after.

  • Which version did this data come out of? Are the system version and the model version written into the log together?
  • Was the procedure settled by a person or by a model? If by a model, do the literature it consulted and the grounds for its judgment survive?
  • How many errors came up during the run, and in the ones recovered automatically, what did the system change?
  • Does the same request yield the same procedure a second time? If it does not, at which step do the two diverge?

All four are hard to reconstruct once an experiment is over. As unattended laboratories multiply, the place where data quality gets asked about moves from the output values to the design of the log. Checking which figures came out matters less than looking first at whether the route they came out by is still there.

Editor's Note

Weighing how far the record can be trusted is not an attempt to take anything away from the work. AutoDNA belongs among the projects that write their own limits down as numbers. The sentence recording that a person stepped into six of the fourteen errors did not have to be written at all. The design that sends code through a virtual laboratory first reads as a judgment that what a machine writes has to be checked by a machine again. What leaves something to be desired here is how the result of that check looks from outside.

Thanks for reading this far. The paper is at Nature, Zou's piece is in the News & Views of the same issue, and the system structure and the instrument list are set out on the project page. If you have taken automatically generated experimental records and put them to use, we would like to hear which entries in the log turned out to help.

R

References

Academic Papers

Official Sources