If you follow self-driving labs, meaning labs where an algorithm decides the next experiment and automated equipment carries it out, most of what you have heard in the past two years is the same story: a large language model reads the literature, writes a protocol and generates control code. AutoDNA, published in Nature on November 4 by a team at Shanghai Jiao Tong University, is worth your attention for a different reason. According to the abstract, the system autonomously orchestrated more than 20 types of instruments across more than 11,000 steps to run nucleic acid synthesis, transcription, amplification and sequencing, with humans limited to hardware maintenance such as replenishing reagents and labware.
My read is that the paper's real contribution is the width of the loop: how many kinds of physical instruments an AI can coordinate to carry a wet-lab experiment, meaning work done with tubes, pipettes and reactions rather than on a computer, from start to finish on its own. It does not show that the AI beats human experts, and the human involvement is much heavier than the abstract suggests. Let's go through it step by step.
Writing the protocol was never the bottleneck; turning it into actions across many instruments is
The abstract makes a pointed criticism of earlier autonomous systems: they have mostly been confined to single objectives, well-defined workflows and simple instrumentation. The authors attribute part of this to an add-on architecture, where AI is a task-specific tool rather than something with a global understanding of the experiment.
AutoDNA starts by translating instrument capabilities into a form an AI can call. The supplementary information describes a hardware abstraction layer that wraps each instrument's functions as atomic services, the smallest callable operations such as pipetting, moving a tube, heating or placing a sample on a magnetic rack. Several agents with distinct roles then work together. A Planner Agent breaks down the user's request, a Protocol Agent drafts procedures from the literature, a Hypothesis Agent proposes what to change next, and Code and Hardware Agents translate the plan into instrument commands and handle runtime errors. Every agent runs on Gemini 2.5 Pro, and the literature search step also uses gemini-2.5-flash and text-embedding-004.
The authors are explicit in their summary of contributions that the key multi-agent contribution is not a new general-purpose planning algorithm but the co-design of agent roles, hardware abstractions and runtime execution. That also answers a reviewer who called the multi-agent design fairly routine. So when you read the paper, treat it as a systems engineering result, not a leap in model reasoning.
Compared with the 2020 mobile robot and 2023's Coscientist, what changed is how many instruments and steps the loop covers
Putting AutoDNA next to two earlier Nature papers shows which part of the problem it moves forward.
On July 8, 2020, Burger and colleagues published the mobile robotic chemist in Nature. According to its abstract, a free-roaming robot operated autonomously for 8 days and ran 688 experiments in a 10-variable space, using batched Bayesian optimization, a statistical method that uses past results to estimate which conditions to try next, to find photocatalyst mixtures 6 times more active than the initial formulations. Its idea was to automate the researcher rather than the instruments, and the decisions came from an optimization algorithm, with no language model involved.
On December 20, 2023, Boiko and colleagues published Coscientist in Nature. According to its abstract, it is a GPT-4-driven system that uses tools such as internet and documentation search, code execution and experimental automation to design, plan and perform experiments, demonstrated across 6 tasks including the optimization of palladium-catalysed cross-coupling reactions, and the abstract itself describes the system as semi-autonomous in places. AutoDNA's supplement characterizes Coscientist as showing only single-objective optimization on a single instrument. That is the AutoDNA authors' description, not Coscientist's own.
| Work | Published | How decisions are made | Published scale figures |
|---|---|---|---|
| Burger et al., mobile robotic chemist | July 8, 2020 | Batched Bayesian optimization | 8 days, 688 experiments, 10 variables |
| Boiko et al., Coscientist | December 20, 2023 | GPT-4 with search and code execution tools | 6 tasks |
| Wu et al., AutoDNA | November 4, 2026 | Multiple agents on Gemini 2.5 Pro | 20-plus instrument types, 11,000-plus steps |
AutoDNA's scale figures trace back to the supplement: one complete DNA data storage experiment used 25 instruments of 16 types for 11,382 atomic-service steps over about 203.5 hours, with synthesis alone taking 112.2 hours, or 55.1% of the total. The three studies tackle different tasks in different fields, so the table is not a ranking. It shows the loop widening, from one optimization algorithm driving one robot, to one language model calling tools, to several agents coordinating more than a dozen instrument types for hundreds of hours.
Four demonstrations show the loop works, not that it beats experts
The paper uses four applications to show the loop is usable, and each one proves something narrower than it may first appear.
The first is optimizing enzymatic DNA synthesis. The key metric is stepwise yield, the success rate of adding each base, where small differences compound quickly as strands get longer. The agents converged on a protocol after 11 iterations, 2 of them aborted, and gel analysis gave an average stepwise yield of 97.8%. The authors compare this with 96.82% reported by Li and colleagues in 2025 using the same enzyme. But the state of the art the agent itself found in the first iteration was 98.7%, from Lu and colleagues in 2022, and the task told it to stop once yield fell more than 1% below that. One reviewer also noted that DNAScript had reported 99.7% coupling efficiency, which by the reviewer's calculation makes AutoDNA's error rate more than 6 times worse. The authors replied that the sequence of that enzyme is not public, so it cannot be tested on their platform, and narrowed the paper's comparison to enzymes with published sequences. The accurate claim is that, among publicly available enzymes, autonomous optimization reached the level of manual optimization.
The second is scheduling concurrent requests. According to the supplement, merging four synthesis protocols cut atomic-service calls from 1,964 to 449, and across two independent runs the mean total execution time fell from 1,433.7 minutes to 454.0 minutes. That is the 3.2-fold speedup in the abstract, and it comes from scheduling, not from faster chemistry.
The third is DNA data storage, where the system completed a full write and read cycle and recovered the original digital content exactly. A reviewer was blunt about the authors' claim that this was a critical step toward commercialization: the factors limiting DNA storage are read and write cost and latency, and neither is improved here.
The fourth is cancer mutation detection, where the system identified and validated a KRAS G13D mutation, a change at position 13 of the KRAS protein from glycine to aspartic acid, that had already been detected clinically. This reproduces a known result, and another reviewer said the main results largely replicate or automate human experiments and asked for a genuinely new discovery.
Human help is not minor: one DNA storage run needed 272 manual refills
The abstract says human intervention was limited to hardware maintenance such as replenishing reagents and labware. That scope is accurate, but you should look at the counts.
Supplementary Table S6 lists manual replenishment during each type of run. The 8-sample nucleic acid test needed 7 buffer refills, while siRNA synthesis and the multi-user concurrent run needed none during execution. The full DNA data storage run needed 79 refills of CB buffer, 23 of water, 84 of 200-microlitre pipette tips and 24 of 50-microlitre tips, among others. Adding up every line for that run gives 163 reagent refills and 109 labware refills, 272 in total; that sum is my own calculation from the table, not a figure the paper states. The cancer detection run was on a similar scale, with 74 CB buffer refills and 63 refills of 200-microlitre tips. Reagent preparation before a run and waste removal and cleanup afterwards are also done by people.
A reviewer seized on exactly this in the second round, calling the framing of AutoDNA as autonomous and without human intervention too strong and asking the authors to state the scope of automation up front. The published abstract now says minimal human intervention.
Fault handling also needs people. In their response the authors report 9 types of wet-lab errors occurring 14 times across all experiments, with 8 events covering 3 error types recovered automatically. The rest were mostly lost communication with the robotic arm or sequencer, a badly seated sequencing chip, or an operator spilling reagent into an instrument and causing a power failure, all of which needed a person to restart, reseat or clean up.
The distinction matters. Humans took no part in experiment design, analysis or decisions, and that is AutoDNA's real advance. But humans remain the loop's logistics crew, and the refill frequency is too high for the lab to run unattended overnight for long stretches.
Code and data are public, cost is missing: what to check if you want to follow up
Reproducibility is better than in many comparable papers. The code availability statement points to the GitHub repository IPADS-X/AutoDNA, licensed under Apache 2.0, with the version used in the study archived on Zenodo. The same Zenodo record holds the sequencing data from the enzymatic synthesis and DNA storage experiments and the agents' outputs, and further data are available from the corresponding authors on reasonable request.
Cost is a blank. I found no hardware purchase cost, reagent and consumable spending, or Gemini 2.5 Pro API spending in the abstract, the supplementary information or the peer review file. The paper's mention of saving time and labour costs is qualitative. Given that one DNA storage run took about 203.5 hours on 25 instruments, that gap matters for judging the authors' science-as-a-service vision.
If you work on lab automation, read this paper as an engineering reference for multi-instrument wet-lab loops. Focus on how the hardware abstraction layer defines atomic services and how errors are tiered for recovery, both of which you can check in the open-source code. Don't read it as evidence that AI scientists can now do research on their own: it reproduces known results, humans set the ceiling for optimization, and humans still run logistics. The next questions are whether refills and fault recovery can be handed to robots too, and how much one of these loops actually costs to run.