An artificial intelligence system called AutoDNA planned and ran DNA experiments across more than 11,000 steps with almost no human help, according to a paper published in Nature on 4 November 2026. Researchers at Shanghai Jiao Tong University built the system as an AI-native laboratory, with seven specialised agents dividing the work between reading the literature, designing protocols, checking reagents, writing the code that drives the hardware, watching the instruments, planning tasks and forming hypotheses.

AutoDNA coordinates more than 20 types of instrument to carry out nucleic-acid synthesis, transcription, amplification and sequencing. A human operator gives it a typed instruction, and the agents then design and run the experiment. Human intervention is limited to hardware maintenance, such as refilling reagents, replacing labware and servicing machines.

Four tests, including DNA data storage

The paper demonstrates the system on four tasks. In enzymatic DNA synthesis, the agents repeatedly adjusted reaction times, reagent concentrations and wash steps until they reached a 97.8 percent stepwise yield, on par with expert human optimisation. Running several protocols at the same time was 3.2 times faster than running them one after another.

In the longest test, the system stored the sentence "Intelligence is the ability to adapt to change" in DNA, encoding it into 78 strands, synthesising and storing them, then sequencing them back. The job took about 203.5 hours and 11,382 hardware steps, and the sentence came back exactly, without error.

In a diagnostic-style test, AutoDNA started with a pre-processed colorectal-cancer sample, designed and synthesised primers, then correctly detected and validated the clinically observed KRAS G13D point mutation.

Not fully autonomous yet

The lab is not fully autonomous, and the paper is open about that. Fourteen hardware errors occurred during the experiments: the agents resolved eight of them, and six needed a person. The team also warns that the randomness built into large language models can change whether the same plan produces the same result when it is run again, a reproducibility problem to watch before such a lab is trusted with high-stakes work.

The tests so far also stay inside nucleic-acid work, so it remains an open question how far the approach extends to other kinds of biology. The system is built around Google's Gemini 2.5 Pro model. The paper, by Wu and colleagues, is titled "An AI-native experimental laboratory for autonomous biomolecular operations". Artificial Science and the AI Frontier Post both reported the Nature paper on 5-6 November 2026.