AutoDNA: AI Agents Ran an 11,382-Step DNA Experiment While Humans Refilled Reagents

This week, in two different laboratories, artificial intelligence stopped suggesting experiments and started doing them. On 4 November, Nature published a paper from Shanghai Jiao Tong University describing AutoDNA, a laboratory in which AI agents plan and run molecular biology experiments on more than twenty machines, with people stepping in mostly to refill reagents and service hardware. The same day, Anthropic reported that its model Claude had helped work out the first three-dimensional structure of a vault particle, a hollow protein shell whose function scientists have not understood since its discovery in 1986. The two announcements are different in almost every way — one is a peer-reviewed paper, the other a company blog post about a preprint — but they point at the same shift: the slowest part of science is moving from thinking up the experiment to carrying it out.

AutoDNA's numbers are the kind that make you stop and reread. The system autonomously orchestrates more than 20 types of instruments across more than 11,000 steps, covering the basic operations of working with DNA: making it, copying it into RNA, amplifying it, and reading its sequence. The team tested it on four tasks. It produced DNA of the same quality as human experts. It ran several protocols at the same time, 3.2 times faster than running them one after another. It stored digital data in DNA and read it back without error — 78 strands encoding the sentence "Intelligence is the ability to adapt to change." And it found and confirmed a cancer-linked mutation, KRAS G13D, in a patient sample. The longest single run, the team reports, lasted about 203.5 hours, roughly eight and a half days, and took 11,382 hardware steps. Through all of it, people were, in the team's own description, mainly limited to disposing of waste and replenishing reagents, acting when the system prompted them.

Anthropic's vault story is smaller in scale and stranger in substance. Vaults are hollow protein shells found in nearly all the cells in our body, about three times the mass of a ribosome, and nobody knows what they do. Under the team's direction, Claude searched a database of about 15 billion predicted proteins, designed the DNA for the candidates, suggested a faster way to isolate the particles, and ran the software that turns microscope images into a 3D model. The result is the first structure of a vault built from a bacteriophage protein: 96 copies of a single protein called L05, enclosing about 25 percent more space than the 78-copy human version. A purification that conventionally takes three days took about 90 minutes; a first structural model was ready 34 hours later, where the team estimates it would normally have taken weeks. The whole project ran under four months.

I am struck by how honest both teams are about what still fails. AutoDNA recorded 14 hardware errors across its experiments; the system fixed eight on its own and six needed a person. The team warns that the randomness built into large language models can change whether the same plan gives the same result twice, and the system cannot yet run experiments in live animals. Claude's mistakes cost real time: one protocol it wrote called for a growth-medium additive that blocked part of the work and set the team back about two weeks, and early in the microscopy analysis it skipped a standard step and produced a flawed map blended with its own mirror image, until the team told it to stop improvising — "Please let's do normal things unless they don't work."

James Zou, a data scientist at Stanford who wrote the commentary Nature ran alongside the AutoDNA paper, put the stakes plainly: an AI system might design a promising biologically active molecule in minutes, but testing it can take weeks or months, and the Shanghai work "can begin to close the gap between proposing an experiment and doing it." Milad Abolhasani, a chemical engineer at North Carolina State University, called AutoDNA remarkable for its degree of autonomy and said the future of science is going to involve a lot of autonomous labs. Both are right, and both understate the question that now matters most. A demonstration is not a discipline. AutoDNA first appeared as a preprint in July 2025; it took about a year of review to reach Nature. The test of this week is not whether one lab can run 11,382 steps — it is whether a second lab, without the original authors in the room, can repeat the trick, and whether the results hold up when it does. That is the line between a striking paper and a new way of doing science, and the next year will show which side of it we are on.