CAi Copilot Streamlines Molecular Design with Agentic Workflows

Early-stage molecular design is a complex, iterative process involving numerous steps from defining broad goals to assessing properties and gathering evidence. While AI methods exist for generating molecules, optimizing goals, and predicting properties, these functions are often fragmented across specialized tools, requiring significant expert coordination. This research addresses this challenge by introducing CAi Copilot, an expert-oriented agent designed to streamline the molecular design workflow. CAi Copilot operates with three integrated layers: a Research Interface Layer to translate intent into executable plans, an Agent Reasoning Layer to guide runs based on interim results, and an Execution Substrate providing molecular tools and services. This system transforms broad molecular design intentions into transparent, traceable workflows that connect interim decisions to candidate-level evidence. Across 45 tasks, CAi Copilot demonstrated superior performance, significantly outperforming other methods in coordinating generation, screening, and multi-criteria evaluation, though limitations in long-horizon execution were noted.