AutoDNA ran 11,382 laboratory steps—but six hardware failures still needed people

The story

An artificial-intelligence system at Shanghai Jiao Tong University has planned and executed long molecular-biology workflows across more than 20 types of laboratory instrument. The system, called AutoDNA, links specialized software agents to robotic handling, DNA synthesis, amplification and sequencing equipment. Its strongest demonstration ran for 203.5 hours and completed 11,382 automated hardware steps to write a sentence into DNA, sequence it and recover the original text exactly.

The peer-reviewed Nature paper describes an ‘AI-native’ laboratory rather than conventional automation with a chatbot added on top. AutoDNA breaks a request into literature retrieval, reagent selection, protocol design, code generation, hardware execution, result analysis and iteration. Each instrument exposes basic capabilities as standardized ‘atomic services’. Agents combine those services into executable workflows while scheduling shared equipment and respecting timing constraints.

That architecture addresses a genuine laboratory bottleneck. AI models can suggest proteins, molecules or experimental conditions quickly, but physical validation still depends on liquid handling, incubation, measurement and troubleshooting. Instruments from different vendors often have incompatible interfaces. By giving agents a common description of what each machine can do, AutoDNA attempts to close the gap between proposing an experiment and performing it.

The researchers tested four distinct capabilities. AutoDNA optimized an enzymatic DNA-synthesis protocol until its reported yield matched an expert-tuned process. It combined compatible work from four synthesis protocols, reducing instrument-service calls from 1,964 to 449 and cutting mean execution time from 1,433.7 minutes to 454.0 minutes—a 3.2-fold improvement over sequential execution on the same platform. That is a scheduling comparison, not evidence that the chemistry itself became three times faster.

The DNA-storage run showed endurance and coordination. The system divided a sentence into 78 DNA strands, synthesized the material, stored it, sequenced it and decoded the result. Exact recovery establishes that this one end-to-end pipeline worked. It does not provide a general bit-error rate across many messages, storage periods or laboratory environments, and the demonstration depended on a heavily engineered instrument stack.

A fourth experiment started with a preprocessed colorectal-tumour sample whose sequence was not provided to the agents. AutoDNA identified a KRAS G13D mutation that clinical testing had already found, designed primers and used them to distinguish the mutant sequence. The result is a workflow proof, not a clinical trial. It says nothing about sensitivity, specificity or performance across patients and should not be described as a new diagnostic test.

The word ‘autonomous’ also needs qualification. People chose the research tasks, configured the physical platform, prepared reagents and equipment, replenished consumables and removed waste. Of 14 hardware errors recorded during the study, the system recovered automatically from eight; six required human intervention. The paper’s supplementary videos show recovery from pipette-tip pickup and pipette-malfunction errors, but successful examples do not establish reliability over routine months-long operation.

Safety depends on more than an agent producing plausible code. AutoDNA checks generated programs and uses a digital representation of the laboratory before physical execution. That kind of preflight validation is essential when probabilistic models control expensive equipment and biological material. Wider deployment will require permission boundaries, immutable logs, contamination controls and clear rules for stopping experiments when sensor data, reagents or generated procedures fall outside validated ranges.

INNOVOX analysis: the advance is best understood as a laboratory operating system. Individual robots have pipetted samples for decades; AutoDNA’s contribution is to make diverse instruments composable enough for agents to coordinate a complete request. If that abstraction proves portable, laboratories could share expensive equipment more efficiently and offer standardized experimental capacity remotely. If every deployment still requires extensive custom integration, the approach will remain a sophisticated facility rather than scalable science-as-a-service.

The next evidence should come from outside the original laboratory. Independent teams need to reproduce the four workflows, connect equipment that the authors did not configure and publish intervention rates rather than only completed tasks. More demanding tests would ask the system to pursue an open scientific question, recognize when evidence is inadequate and revise its plan without a known answer. AutoDNA demonstrates that AI agents can cross from digital plans into wet-lab operations; it does not yet show that they can independently conduct reliable science.

INNOVOX analysis

AutoDNA’s consequential innovation is the software and hardware abstraction that lets agents translate experimental intent into coordinated physical operations. The reported biological tasks are demonstrations of that orchestration layer, not discoveries. The system moves human work from repeated bench procedures toward platform engineering, validation, maintenance and exception handling rather than removing scientists from the loop.

What to watch

Watch for independent multi-site replication, operation on instruments not integrated by the original team, published failure and intervention rates, and prospective experiments whose answers are not already known. Clinical or hazardous work will also require auditable safety controls beyond successful code generation.