# Flow: Modularized Agentic Workflow Automation

**Authors:** Boye Niu, Yiliao Song, Kai Lian, Yifan Shen, Yu Yao, Kun Zhang, Tongliang Liu  
**arXiv:** [2501.07834](https://arxiv.org/abs/2501.07834)  
**Date:** 2025-01-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.07834v2/

## Key points

- Effective adjustment of agentic workflows during execution is critical for handling unforeseen challenges and changing conditions
- Workflows defined as activity-on-vertex (AOV) graphs enable continuous refinement through dynamic subtask allocation adjustments based on historical performance and previous AOVs
- Modularity in workflow design is emphasized to enhance framework performance, evaluating parallelism and dependency complexity
- Empirical results across various practical tasks show significant improvements in efficiency through dynamic workflow refinement and modularization
- Flow framework demonstrated an 80% success rate in generating HTML websites with detailed sections and interactive features like maps and tables

## Summary

, , , , 

In the realm of multi-agent frameworks powered by large language models (LLMs), the effective adjustment of agentic workflows during execution is a critical yet underexplored area. The ability to adapt workflows in real-time is essential for handling unforeseen challenges and changing conditions, ensuring the efficient completion of complex tasks. In this study, workflows are defined as activity-on-vertex (AOV) graphs, enabling continuous refinement through dynamic subtask allocation adjustments based on historical performance and previous AOVs. To enhance framework performance, modularity in workflow design is emphasized, evaluating parallelism and dependency complexity. This approach allows for efficient concurrent execution of subtasks, effective goal achievement, and enhanced error tolerance within the proposed multi-agent framework. Empirical results across various practical tasks showcase significant improvements in efficiency through dynamic workflow refinement and modularization. The study compares four frameworks – AutoGen, MetaGPT, CAMEL, and Flow – across tasks such as website design and Gobang game development. AutoGen successfully generated outputs but faced compilation errors and length issues. MetaGPT excelled in content generation but struggled with page count requirements. CAMEL produced valid outputs but fell short on page count specifications. Flow consistently generated compilable outputs with one instance of repetitive content. In website design tasks specifically, Flow demonstrated an 80% success rate in generating HTML websites with detailed sections and interactive features like maps and tables. The framework's update process was illustrated through a conference website creation example, showcasing its ability to identify redundancies and optimize workflow efficiency. Overall, the study highlights the importance of dynamic workflow adjustment in multi-agent frameworks powered by LLMs for efficient task execution in real-world scenarios. The emphasis on modularity and continuous refinement proves instrumental in achieving optimal performance across diverse tasks.