--- title: "Chen Haibo's lab ran 11,382 hardware steps from one sentence" date: 2026-11-09 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-chen-haibo-autodna language: en --- # Chen Haibo's lab ran 11,382 hardware steps from one sentence > Shanghai Jiao Tong's Chen Haibo published AutoDNA in Nature, an AI-native lab that plans and runs biomolecular experiments across 20-plus instruments. The request was a sentence, typed the way you would ask a colleague: read what was written into this DNA, and tell me what it says. What followed was not a person walking to a bench. A planner agent broke the request into tasks, a reagent agent checked what was on hand, a code agent wrote instrument commands, a hardware agent ran them, and an analysis agent read the output and sent corrections back. Across the full write-and-read cycle the system executed 11,382 hardware operations and recovered the original text without a single error. The work was published in Nature on 4 November. ## Key takeaways - Chen Haibo (陈海波), Wang Zhaoguo (王肇国) and colleagues at Shanghai Jiao Tong University (上海交通大学), with Fan Chunhai (樊春海) and Wang Fei (王飞), published **AutoDNA**, an AI-native autonomous laboratory, in *Nature* on **4 November 2026**. - The system orchestrates **more than 20 types of instrument** and has run **more than 11,000 hardware steps** in a single project, with human intervention limited to restocking reagents and maintaining hardware. - It uses a planner plus six specialist agents and a hardware abstraction layer that exposes each instrument as a callable "atomic service". - Reported results include **97.8% single-step yield** in enzymatic DNA synthesis, a **3.2-fold** gain in parallel protocol throughput, and detection of a **KRAS G13D** mutation in a colorectal cancer sample whose sequence was unknown to the system. - *Nature* ran the paper as a front-page story and commissioned a News & Views commentary from Stanford's James Zou. ## The add-on problem Autonomous labs are not new. Chemistry and materials groups have run closed-loop systems for years: a robot dispenses, an analyser measures, an optimiser picks the next condition. Those systems share a design assumption. A human decides which instruments exist, what the workflow is, and which knobs the algorithm may turn. The AI is bolted on to a process that was specified without it — what the SJTU team calls an "add-on" architecture. That works when the experiment is one clean loop on simple equipment. It breaks in molecular biology, where a single protocol can touch twenty devices, where timing and temperature matter to the minute, and where the next step depends on what the last measurement actually showed. The team's argument is that you cannot bolt autonomy onto a workflow that was designed to be operated by hands. ## Instruments as callable services The fix is familiar to anyone who has worked on operating systems: build an abstraction layer. AutoDNA wraps each instrument's capabilities — pipetting, incubation, thermal cycling, magnetic separation, sequencing — in a description a model can read, plus a unified code interface. The model does not need to know that two liquid handlers have different command languages. It needs to know that this one can transfer 20 microlitres to a given well. Above that layer sits the planning stack: a central planner that decomposes the request, and specialist agents for literature, reagents, protocol, code, hardware and hypothesis. They hand work to each other and revise when a result comes back wrong. Underneath sits hardware capable of micron-level positioning and microlitre-scale liquid handling, with a scheduler that tracks instrument state and enforces timing constraints. The control terminals in the build run on OpenHarmony. ## Four demonstrations, one of them clinical The paper reports four applications, and the details are more interesting than the headline. In enzymatic DNA synthesis, the system optimised for yield and time together, reaching 97.8% single-step yield — the team says comparable to what an expert achieves by hand. In parallel scheduling, it merged compatible operations and reused resources, improving throughput 3.2-fold. In DNA data storage, it took a user's text, encoded it, synthesised the DNA, stored it, sequenced it with a nanopore device and read the sequence back, recovering the original message exactly. The fourth is the one clinicians will read. Given a colorectal cancer sample with unknown sequence, AutoDNA sequenced it, identified mutations, designed primers and validated them; the first primer set it designed confirmed a KRAS G13D mutation that clinical testing had already flagged. That is a small result and a meaningful one: the system did not merely execute a protocol, it closed a diagnostic loop. ## Why this is a systems result Chen works in SJTU's IPADS lab, a group known for systems research, and the paper's contributions are recognisably systems contributions: an abstraction, a scheduler, fault handling, multi-user resource management. The biology is the workload. ## What it does not prove AutoDNA is not a general scientist. It does not decide which disease matters, and its hypothesis agent operates inside a protocol space that people defined. The 97.8% figure is a single-step synthesis yield, not an end-to-end success rate, and "zero error" refers to one short message in a DNA storage cycle, not to error-free biology. Nor is this the first autonomous lab. Groups in the UK, the US and elsewhere have run closed-loop chemistry for years; what is new here is the breadth of instrument coordination and the explicit claim that the AI should own the workflow rather than assist it. The honest summary is that the paper demonstrates a credible architecture and four convincing demos, in one lab, with one team's instruments. ## The cost question nobody has answered A system like this is capital-intensive: twenty instruments, a robotic arm, environmental control and a software team to keep the drivers alive. The paper reports capability, not cost per experiment. That matters because the economic case for autonomous labs is not speed on a single protocol. It is utilisation — running experiments overnight, in parallel, without a trained operator in the room. Whether AutoDNA's uptime and maintenance burden make that case is an operational question the publication does not settle. According to Li An, chief scientist at BrainNet (脑机网), the decisive metric for autonomous labs will be how many protocols a system can absorb without a human rewriting its abstractions. ## Honest limitations The primary source is the Nature paper, published 4 November 2026 as article s41586-026-11228-6, plus SJTU's own release and a ScienceNet (科学网) report; I read the abstracts, institutional summaries and figures reproduced in those releases, not the full methods section or supplementary code. The 11,382-step figure refers specifically to the DNA storage read/write cycle, while "more than 11,000 steps" describes the broader project; the two should not be conflated. The 97.8% yield, 3.2-fold throughput gain and KRAS G13D result are the authors' reported measurements on their own platform and have not been independently replicated. "Zero error" applies to a short encoded message. Human intervention was limited but not zero: reagents, consumables and hardware maintenance were still manual. I did not verify the instrument count beyond the team's own description. ## What readers can do now - Read the methods before the press release: the interesting questions about an autonomous lab are instrument coverage, failure handling and uptime, not the demo count. - If you run a lab, inventory how many of your protocols could be expressed as callable instrument services — that number, not the model, sets the ceiling. - Treat single-step yields as single-step yields. Ask for end-to-end success rates per completed experiment. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-chen-haibo-autodna Free to quote with attribution and a link to the original.