An AI-native experimental laboratory for autonomous biomolecular operations

Researchers at Shanghai Jiao Tong University have built AutoDNA, a laboratory where AI agents design and carry out DNA experiments from a single natural-language prompt, with humans mostly restocking reagents and fixing hardware. The system, described in a Nature paper dated November 4, 2026, uses a team of large-language-model agents coordinated by an AI "coordinator" that controls robotic arms operating over 20 types of lab equipment. In testing, AutoDNA completed an entire DNA read-write cycle in around 200 hours with precision comparable to expert scientists, according to SingularityHub. The most demanding test encoded and decoded a message in DNA, requiring 11,382 hardware steps over roughly eight days (203 hours), during which all 78 synthesized DNA strands were successfully recovered even with a few missing letters. The system also optimized a polymerase chain reaction process to a yield of nearly 98 percent, matching human experts, and identified a clinically significant KRAS G13D mutation in a colorectal cancer sample. Humans intervened periodically to replenish pipette tips, glass tubes, and reagents, and to empty trash; the system paused and alerted technicians when supplies ran low. Across experiments, AutoDNA logged 14 hardware errors, six of which required a person to fix. The lab cannot yet handle delicate experiments involving living cells or animals. James Zou of Stanford University, who was not involved, wrote that the work shows how AI systems can begin to close the gap between proposing an experiment and doing it, and noted the planner agent demonstrated a sort of common sense by capping test tubes to prevent evaporation without explicit instruction. Reshma Shetty, co-founder of Ginkgo Bioworks, told Nature the system is "a teaser of what might be possible in the future." A preprint version of the work was posted on arXiv in July 2025. The researchers, led by Haibo Chen with Mingyu Wu, Zhaoguo Wang, and Jiabin Wang, describe the approach as AI-native, built around multiple cooperating agents rather than a single model. Safety concerns include the potential for self-driving labs to synthesize DNA encoding dangerous proteins; the team suggests built-in screening software, access controls, and experiment ledgers as safeguards. Digital twins of the lab are proposed as virtual sandboxes for agents to rehearse and catch errors before real-world runs.