Skip to main content
Back to timeline
arXivSource publication:

Chemists write only the target, and VibeSyn's agents run a ten-step total synthesis: active human input is 7.3%

Synopsis

A team led by Zhejiang University built VibeSyn, a hierarchical agent system that turns a chemist's molecular objective into synthesis decisions revised as experimental evidence accumulates, then drives computational tools and automated devices to execute them. In a ten-step catharanthine total synthesis the system recorded 722 actions, of which agent reasoning was 69.7% and tool or device outcomes 18.8%, while active human input and human confirmation were 7.3% and 4.2%; starting from 29.8 mmol of 3-ethylpyridine it produced 160 mg of product. Across three campaigns synthesizing cyclic peptides containing unnatural amino acids, active human input and confirmation were 7.5% and 5.2%.

Source-provided article image: Vibe synthesis: hierarchical agentic decision-making for intent-driven multistep molecule synthesis
arXiv

Interpretation

Multistep synthesis no longer follows a pre-written workflow; agents decide from the evidence accumulated at each step whether to proceed, correct a failure, optimize, or return upstream and change the route. Earlier computer-aided route design and automation platforms were typically connected through predefined tasks and hard-coded workflows, so new knowledge or context-dependent judgments required additional coding to enter the process. The ten-step catharanthine campaign recorded 722 actions: 503 agent reasoning records, 136 tool or device outcomes, 53 active human inputs and 30 human confirmations, yielding 160 mg of product from 29.8 mmol of 3-ethylpyridine.

One manager agent splits the overall objective into chemically bounded tasks and assigns them to function-specific agents, which call tools or devices and write results back into an evolving synthesis state. The manager agent maintains the objective and accumulated evidence, while task agents invoke computational tools and automated devices through structured request-response interfaces and decide whether further action is needed before reporting. The system comprises 10 task agents connected to 10 computational and multimodal analysis tools and 7 automated synthesis devices, spanning route and condition design, reaction development, downstream processing, product verification and peptide synthesis; task agents primarily use OpenAI o3 (o3-2025-04-16).

Evidence can overturn an upstream design: when one pyrazole monomer route gave only the wrong regioisomer in three trials, the manager agent traced the failure to route design and reassigned it, and a milder six-step route delivered the target monomer. The original four-step route began with a Knorr condensation; three trials varying temperature and solvent gave 0% of the desired product and consistently formed the undesired 3-CF3/5-phenyl regioisomer. The redesigned six-step beta-methoxy-enone route reached the desired pyrazole through a [3+2] cyclocondensation and dehydration, affording 1.89 g of crude Fmoc-protected monomer in 15.4% overall six-step yield.

Perspective

The demonstrated scope is three settings: evidence-to-action decision tasks, a ten-step catharanthine total synthesis, and structure-to-compound synthesis of cyclic peptides containing unnatural amino acids; the intended readers are chemistry and engineering researchers working on automated synthesis and agent decision-making. Current capability is bounded by configured task-agent expertise and the operating ranges of connected tools and devices, and steps beyond the automated hardware's capability, such as in situ generation of trimethylsilyl iodide, are assigned to a human chemist.

Decision reliability depends on evidence quality and on how faithfully the campaign state represents the experiment; when evidence is ambiguous the system does not yet explicitly represent uncertainty to trigger additional measurements or targeted human input. Experimental state currently comes from isolated images and analytical traces rather than continuous multimodal evidence from time-resolved video, in situ spectroscopy and equipment data.