AutoDNA connects AI agents to physical biology equipment, allowing the system to plan and carry out workflows that make, copy and read DNA. In a Nature paper published on 4 November, researchers report four demonstrations: optimizing DNA synthesis, coordinating experiments that share equipment, storing digital information in DNA and identifying a known cancer-associated mutation.

The work shows how AI-generated plans can become physical laboratory operations through a purpose-built automation system. Its results depend on the equipment, workflows and controls used in those demonstrations. People still replenish supplies and maintain the hardware.

The authors report orchestrating more than 20 types of instruments across more than 11,000 steps. Those counts describe instrument categories and operations, rather than independent experiments. The accessible abstract does not define how the steps were counted or the period they cover.

From agent plans to instrument actions

The system has several layers. According to the versioned AutoDNA repository, a web interface connects to a multi-agent engine, with specialized roles for tasks including literature searches, protocols, reagents, code and hardware. A scheduler arranges equipment use and coordinates workflows. Firmware then translates scheduling commands into actions by physical instruments.

That middle layer matters when two experiments need the same heater or liquid-handling equipment. A plausible protocol still needs an executable order of operations, available supplies and compatible devices.

The hardware was co-designed with the software. The repository lists equipment such as robot arms, heater shakers, fluorometers and centrifuges, and requires instruments to support the system’s interfaces. It warns that full reproduction is not guaranteed without the hardware. Downloading the code alone therefore does not recreate the laboratory.

Nature’s accompanying reporting places the team at Shanghai Jiao Tong University in China and describes the broader effort to coordinate complete workflows beyond automation of individual laboratory operations. AutoDNA’s four demonstrations give that ambition a concrete, testable form.

DNA synthesis: improving short sequences against an expert baseline

The first demonstration used enzymes to build short DNA sequences. AutoDNA iteratively adjusted the process, with the authors reporting synthesis yields that matched an expert-optimized approach.

The public Figure 3 caption reports four separately synthesized samples per optimization iteration. A gel analysis, which separates DNA pieces by size, was independently repeated twice with similar results and used known-length DNA as size references. Sequencing provided another way to examine synthesis yield and errors.

The reported optimization applies to the short sequences tested. The comparison with expert-optimized yields remains the authors’ finding because full baseline conditions and optimization methods were unavailable in the material reviewed.

Scheduling: sharing equipment reduced execution time

The second demonstration addressed a familiar automation problem: several jobs competing for the same resources. AutoDNA consolidated instrument-control programs and resolved conflicts over equipment use, including allocation of a heater and a thermal cycler, which changes temperature during DNA-processing reactions.

The authors report reduced execution time through scheduling shared equipment. Figure 4 compares optimized scheduling with sequential execution for multiple DNA-synthesis protocols and a multi-user scenario; its timing and instrument-utilization panels report two independent experiments.

The timing result applies to execution of the demonstrated workflows on shared laboratory equipment.

DNA storage: recovering digital content from 78 encoded strands

For the third demonstration, AutoDNA carried out a DNA data-storage workflow. Digital information was encoded into DNA sequences, which were synthesized, purified and stored. Reading involved sequencing the DNA and decoding the information.

The authors report exact recovery of the original digital content in the demonstrated cycle. The Figure 5 caption describes 78 distinct data-encoded DNA strands that were synthesized and sequenced. That is a count of strands, not 78 independent end-to-end storage trials; the caption does not state how many complete write-and-read cycles were independently repeated.

Exact recovery of the decoded information also does not mean the DNA chemistry was error-free. The figure separately examines sequencing yield, errors and how bit recovery changes with sequencing coverage, the amount of sequence-reading data collected.

KRAS testing: identifying a variant and checking it against controls

The fourth demonstration combined sequencing, analysis and a targeted follow-up assay. The researchers report identifying a previously clinically detected KRAS G13D variant, then checking it using allele-specific PCR. PCR copies selected DNA; an allele-specific assay is designed to favour a particular genetic variant.

The Figure 6 caption describes a mutant target, a wild-type control from one healthy donor and a no-template control, which checks what happens without target DNA. The targeted assay reports three independent experiments. These are experimental repeats, not three patients or a clinical cohort.

The team also examined primers made by AutoDNA. Primers are short DNA pieces used to start copying a sequence. Digestion products from commercially synthesized primers supplied size references in a gel experiment, independently repeated three times with similar results.

Together, these results support mutation identification and targeted confirmation. The accessible caption does not establish a patient cohort size, diagnostic sensitivity or specificity, prospective clinical validation or treatment benefit.

People still replenish and service the laboratory

The paper describes human intervention as limited to hardware maintenance, including replenishing reagents and laboratory supplies. The DNA-storage figure is more specific: intervention during writing was primarily reagent and pipette-tip replenishment. Nature’s reporting also describes people refreshing reagents and servicing hardware.

The repository carries a separate warning that users running some examples may occasionally need to enter input through a command-line interface. That warning concerns the published examples; it does not establish that those interventions occurred during the paper’s experimental runs.

These dependencies are part of what it takes to operate the system. A laboratory adopting this approach would need compatible instruments, integrated control software and people to keep the physical setup supplied and working.

The Antigravity chemistry demonstration shows an earlier stage of a research workflow, where an agent proposes molecules and routes for researchers to test.

The evidence reviewed here comprises the paper’s accessible abstract and figure captions, public repository material and accessible accompanying reporting. The subscription-limited main methods and underlying supplementary files were not inspected, and the experiments were not independently reproduced for this article. For a lab assessing whether it could repeat the work, the detailed methods, equipment interfaces and servicing requirements would be essential next checks.