FLORA: GNNs as Predictors of Agentic Workflow Performances

Yuanshuo Zhang, Yuchen Hou, Bohan Tang, Shuo Chen, Muhan Zhang, Xiaowen Dong, Siheng Chen

Keywords: Graph Neural Networks, Multi-agent Systems, Agentic Workflow Optimization

Abstract: Agentic workflows invoked by Large Language Models (LLMs) have achieved remarkable success in handling complex tasks. However, optimizing such workflows is costly and inefficient due to extensive invocations of LLMs. To fill this gap, this paper formulates agentic workflows as computational graphs and proposes **FLORA** (work**FLO**w g**RA**ph neural networks), which are variants of Graph Neural Networks (GNNs), as efficient predictors of agentic workflow performances, thus avoiding repeated LLM invocations for evaluation. To empirically ground the effectiveness and efficiency of FLORA, we construct **FLORA-Bench**, a unified platform for training and benchmarking predictors of agentic workflow performances. With extensive experiments, we arrive at the following conclusions: (1) FLORA is a simple yet effective method to accurately predict agentic workflow performances, and (2) it demonstrates significant practical benefits, achieving up to a 125× speedup with minimal performance loss. These conclusions support new applications of GNNs and a novel direction towards automating agentic workflow optimization. All codes, models, and data will be public.

Submission Type: Full paper proceedings track submission (max 9 main pages).

Submission Number: 21