Preprint: VibeSyn AI system adaptively executed a 10-step synthesis of catharanthine

·

A Zhejiang University-led team says in a new arXiv preprint that it built an AI-directed chemistry system capable of adaptively carrying out long, multistep synthesis campaigns, including a 10-step synthesis of the natural product catharanthine, with relatively limited human intervention.

The paper, titled “Vibe synthesis: hierarchical agentic decision-making for intent-driven multistep molecule synthesis,” was submitted to arXiv on Nov. 12, 2026, as arXiv:2611.18772v1. It is a preprint, meaning the results described by Jie Chen and colleagues have not been peer reviewed.

That distinction matters because multistep chemical synthesis is not just a matter of following a fixed recipe. Chemists often have to revise what happens next as experimental results come in from earlier steps. Much of the earlier AI work in chemistry has focused on planning routes backward from a target molecule, known as retrosynthesis, or on optimizing individual reactions in closed-loop systems. The new paper positions its system, called VibeSyn, as a step toward adaptive execution across many chemically dependent steps.

The authors describe VibeSyn as an “intent-driven” hierarchical agent system that takes a chemist’s target molecule and uses accumulating experimental evidence to decide what to do next through software tools and lab automation. The implementation described in the paper includes a Manager Agent coordinating 10 task agents, 10 computational or multimodal tools, and seven automated synthesis devices. The report also says some steps still required “safety-bounded human participation” when work fell outside the configured automation envelope.

The strongest evidence in the preprint centers on the catharanthine campaign. According to the paper, VibeSyn executed a route based on 1987 literature, starting from 29.8 millimoles of 3-ethylpyridine and ultimately producing 160 milligrams of catharanthine. The authors say the 10-step campaign generated 722 recorded workflow actions: 503 agent reasoning records, or 69.7%; 136 tool or device outcomes, or 18.8%; 53 active human inputs, or 7.3%; and 30 human confirmations, or 4.2%.

More broadly, the paper says that across its long-horizon campaigns, active human input accounted for no more than 7.5% of recorded actions, while simple human confirmation accounted for no more than 5.2%. Those figures are notable because they suggest the system was not fully autonomous, but also was not being manually directed at every decision point.

The preprint reports two other demonstrations as well. One involved what the authors call evidence-grounded experimental decision-making on smaller tasks. The other was a structure-to-compound synthesis workflow for cyclic peptides containing unnatural amino acids, a test meant to show the system could translate a molecular design into a real synthesized compound in a more specialized setting.

The work involves researchers from Zhejiang University and collaborators from Tsinghua University, ChemLex Technology Co., Suzhou Novartis Technical Development Co., and Harvard Medical School.

The authors say the code is publicly available through GitHub and Zenodo. But the public release is not a turnkey replication package. According to the repository README described in the paper, the distributed Docker image does not include the model services, databases or laboratory devices needed for full operation. That means outside researchers can inspect parts of the system, but full end-to-end replication is not immediately available from the posted materials alone.

For now, the paper should be read as an unreviewed research claim rather than a settled advance. Still, it points to an important technical goal in automated chemistry: systems that can change course as evidence accumulates instead of rigidly following a preset workflow. As the authors write in the abstract, they “propose and demonstrate a vibe synthesis system (VibeSyn), an intent-driven architecture in which hierarchical agents translate chemist-defined molecular objectives and accumulating evidence into adaptive decisions executed through computational tools and laboratory automation.”

Tags: #ai, #chemistry, #automation, #synthesis