AutoDNA Coordinated 20 Lab Instruments Through 11,382 Operations

A laboratory at Shanghai Jiao Tong University has connected AI agents, scheduling software and automated hardware into one system that can plan and execute long biomolecular workflows. The AutoDNA platform coordinated more than 20 types of instruments and completed experiments spanning DNA synthesis, data storage and mutation detection, according to a peer-reviewed Nature paper published on November 4.

Key number: 11,382 hardware operations. AutoDNA executed that many physical actions during an end-to-end DNA data-storage cycle and recovered the encoded information exactly.

Turning a request into instrument commands

The system uses multiple specialized agents to divide a task into literature review, reagent selection, protocol design, code generation, hardware execution, result analysis and iterative adjustment. Rather than exposing every device through a different control scheme, the team represented instrument capabilities as “atomic services” described in natural language and called through common software interfaces.

That abstraction lets the planning layer combine operations from liquid handlers, incubators, sequencers and other equipment without writing every workflow from scratch. A separate execution layer checks programs, tracks instrument state and schedules shared resources while the hardware handles micrometre-scale positioning and microlitre-scale liquids.

Four tests of the same architecture

The researchers evaluated the platform in four applications. It matched expert-optimized yields for enzymatic DNA synthesis, ran concurrent protocols 3.2 times faster than a sequential baseline, completed an end-to-end DNA storage workflow and identified a KRAS G13D mutation that had already been detected clinically.

The DNA storage demonstration was the longest orchestration test. AutoDNA linked encoding, DNA synthesis, preservation, sequencing and decoding, coordinating computational systems with laboratory instruments. Exact recovery showed that the full chain completed successfully, but it does not by itself measure commercial storage cost, long-term media durability or performance at industrial scale.

The mutation result also needs careful wording. The platform did not discover a new cancer marker or diagnose patients prospectively. It reproduced the detection of a known clinically observed variant, which tests workflow integration rather than clinical accuracy in a population.

From digital feedback to physical execution

NewTqnia previously covered AI agents that revised molecular designs after real laboratory failures. That work demonstrated feedback between computation and supervised experiments. AutoDNA pushes the integration further by giving the planning system a unified way to call physical equipment through a long workflow.

A second comparison is AI-guided X-ray imaging that redirected a beamline during measurement. Both systems close a loop between analysis and instruments, but AutoDNA addresses multi-device biomolecular protocols rather than adaptive imaging at one scientific facility.

Why co-design matters

The authors describe AutoDNA as AI-native because the models, protocols and instruments were designed to work together. That differs from adding a chatbot to an existing laboratory whose devices expose inconsistent controls and whose procedures assume constant human intervention. It is a meaningful engineering distinction, although it also makes transfer to an ordinary laboratory harder.

The paper reports that the software and data are available, including archived code, which gives other researchers a basis for inspection. Reproducing the full platform would still require compatible hardware, device interfaces, safety procedures and local validation. Code availability is not the same as a turnkey replica.

What would show generality?

The decisive next test is replication outside the originating laboratory. Researchers would need to connect different instrument brands, run protocols the system was not co-designed around and publish failure rates, recovery behavior and human-intervention time. If those tests succeed, the important advance will be less about replacing scientists than about making complex experimental execution programmable, inspectable and repeatable.