Vibe synthesis: hierarchical agentic decision-making for intent-driven multistep molecule synthesis

Jie Chen, Guikun Chen, Feng Wan, Ning Xu, Wen Liu, Yixiang Ruan, Yiyi Qiu, Haohua Li, Wei Wang, Yi Wang, Xiaoxue Wang, Peng Wang, Jiazhou Zhu, Jun Xuan, Chengxi Li, Jianhong Xu, Wenguan Wang, Yiming MoarXiv 2026 · arXiv

TL;DR

VibeSyn is a hierarchical AI agent system that automates complex, multi-step chemical synthesis by translating high-level "intent" into adaptive, evidence-based actions. By integrating LLM reasoning with real-time multimodal feedback (like image analysis and NMR), the system successfully executed a 10-step total synthesis of catharanthine with less than 8% active human intervention.

What Problem Does This Paper Address?

Traditional laboratory automation is rigid. Most systems rely on "hard-coded" workflows—pre-defined scripts that execute a specific sequence of tasks. However, chemistry is notoriously unpredictable. An intermediate might not dissolve as expected, a reaction might produce an unforeseen byproduct, or a purification step might fail to resolve two compounds. In these cases, a hard-coded robot simply stops, requiring a human expert to diagnose the issue and rewrite the code.

This paper addresses the "long-horizon" challenge: how to build a system that can manage a multi-step synthesis campaign where decisions must be revised on the fly based on accumulating experimental evidence. The authors propose moving from implementation-driven automation to intent-driven "vibe synthesis," where the AI understands the goal and navigates the experimental trajectory autonomously.

Methods

The authors developed VibeSyn, a hierarchical agentic architecture:

  • Manager Agent (MA): The "project lead" that maintains the overall objective, tracks the campaign state, and decomposes the synthesis into chemically bounded tasks.
  • Task Agents: Specialized agents for specific roles, such as Retrosynthesis Design (RDA), Condition Recommendation (CRA), Reaction Optimization (ROA), and Spectrum Analysis (SA).
  • Multimodal Feedback: The system doesn't just process text. It uses GPT-5-based image analysis tools to "see" experimental states—interpreting the color of a solution, the presence of an emulsion during workup, or the spots on a TLC plate.
  • Tool & Hardware Integration: Agents are connected to 10 computational tools (e.g., Bayesian Optimization, NMR predictors) and 7 automated hardware devices (e.g., reactors, liquid handlers, flash chromatography systems) via a unified configuration framework.
  • Reasoning Engine: The system primarily utilizes OpenAI’s o3 model for high-level chemical reasoning and instruction following.

Datasets

The paper does not rely on a single training dataset in the traditional sense. Instead, it leverages:

  • Parametric Knowledge: The internal chemical knowledge of the base LLMs (o3 and GPT-5).
  • RAG (Retrieval-Augmented Generation): A curated library of reference handbooks (e.g., for Bond Dissociation Energies) used to provide quantitative data for decision-making.
  • Experimental Benchmarks: The system was evaluated on three real-world campaigns:
  • Fe-catalyzed C–H functionalization: Testing knowledge-driven condition selection.
  • 10-step Total Synthesis of Catharanthine: A complex, long-horizon campaign based on a 1987 literature route.
  • Cyclic Peptide Library: A structure-to-compound task involving the synthesis of unnatural amino acids (uAAs) and their assembly into 15 non-redundant peptides.

Key Insights and Ideas

  • Evidence-Grounded Transitions: The most striking feature is the system's ability to adapt. For example, in Step 1 of the catharanthine synthesis, the agent "saw" a milky suspension, inferred the product was crashing out, and autonomously instructed the robot to add ethanol to homogenize the solution before transfer.
  • Traceback-Guided Redesign: If a reaction fails to yield the desired product (e.g., forming the wrong regioisomer), the Manager Agent can "trace back" the failure to the route design phase and task the Retrosynthesis agent with finding an alternative pathway.
  • Symmetry Awareness: During the peptide library construction, the Peptide Agent recognized the $C_2$-symmetry of the target scaffold, reducing the required experiments from 25 to 15 non-redundant combinations—a classic "chemist's trick" performed autonomously.
  • Safety and Constraints: Human chemists set "safety boundaries." If a task exceeds the hardware's capability (like handling corrosive iodine reagents), the system identifies the limitation and requests manual intervention while maintaining control over the rest of the workflow.

Key Results

  • Catharanthine Synthesis: Successfully produced 160 mg of the complex natural product over 10 steps. The workflow involved 722 recorded actions, with agent reasoning accounting for 69.7% of the activity.
  • Peptide Library: Synthesized 15 high-purity (>95%) cyclic peptides, coordinating monomer preparation with solid-phase assembly.
  • Minimal Human Load: Across all campaigns, active human input (like submitting tasks or adjusting hardware) was kept below 7.5%, and simple confirmations were below 5.2%.
  • Optimization Success: The system successfully replaced toxic reagents (like HMPA) from old literature procedures with safer, optimized alternatives (like DABCO/MeCN) through iterative Bayesian optimization.

Limitations and Future Work

  • Expertise Bounds: The system is still limited by the "chemical common sense" of the underlying LLM and the specific capabilities of the connected robots.
  • Uncertainty Representation: The authors note that agents should be better at representing uncertainty, knowing exactly when to trigger a new measurement or ask a human for help.
  • From Images to Video: Current multimodal interpretation relies on static images. Future versions could use time-resolved video or in-situ spectroscopy for a more granular view of reaction kinetics.
  • Physical Flexibility: While VibeSyn is modular, it still uses task-specific devices. The next step is integrating more flexible, reconfigurable robotics that can manipulate a wider variety of labware and materials.

Summary updated

Abstract

The ability to construct complex molecules underpins advances in medicine and the molecular sciences. Yet long-horizon multistep synthesis requires decisions to be revised as evidence accumulates across chemically dependent operations, a capability not readily supported by artificial-intelligence tools and automated platforms organized around predefined tasks and hard-coded workflows. Here we 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. We evaluate VibeSyn in three demonstrations of increasing scope and autonomy, progressing from evidence-grounded experimental decisions to adaptive execution of a ten-step total synthesis of catharanthine and structure-to-compound synthesis of cyclic peptides containing unnatural amino acids (uAAs). Across the long-horizon campaigns, active human input and simple confirmation account for less than 7.5% and 5.2%, respectively, of recorded workflow actions and outcomes. VibeSyn points towards a future in which chemists define synthetic intent and agent-directed laboratories autonomously navigate the evolving experimental trajectory required to deliver the target compound.
BibTeX
@misc{chen2026vibe,
  author = {Jie Chen and Guikun Chen and Feng Wan and Ning Xu and Wen Liu and Yixiang Ruan and Yiyi Qiu and Haohua Li and Wei Wang and Yi Wang and Xiaoxue Wang and Peng Wang and Jiazhou Zhu and Jun Xuan and Chengxi Li and Jianhong Xu and Wenguan Wang and Yiming Mo},
  title = {{Vibe synthesis: hierarchical agentic decision-making for intent-driven multistep molecule synthesis}},
  year = {2026},
  eprint = {2611.18772},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18772},
}
Figure