AFlow: Automating Agentic Workflow Generation

Source: arXiv Authors: Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Yuyu Luo, Chenglin Wu Date: 2024-10-14 Primary category: cs.AI All categories: cs.AI, cs.CL, cs.LG, cs.SE

Abstract

AFlow reframes workflow optimization as a search problem over code-represented agent workflows, where LLM-invoking nodes are connected by edges. It explores the workflow space with Monte Carlo Tree Search, iteratively refining workflows through code modification, tree-structured experience, and execution feedback. Across six benchmark datasets, the paper reports an average improvement over prior workflow-generation baselines while also reducing inference cost through better workflow structure rather than only larger models. For this wiki, the paper matters because it treats workflow topology itself as an optimizable artifact.