# A Survey on LLM-Based Agentic Workflows and LLM-Profiled Components

**Authors:** Xinzhe Li  
**arXiv:** [2406.05804](https://arxiv.org/abs/2406.05804)  
**Date:** 2024-06-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2406.05804v1/

## Key points

- Recent advancements in Large Language Models (LLMs) have revolutionized the development of sophisticated agentic workflows.
- LLMs have surpassed traditional single-path prompting techniques such as Chain-of-Thought (CoT).
- The survey focuses on LLM-Profiled Components (LMPCs) within agentic workflows, excluding non-LLM components.
- Generative Large Language Models (GLMs or LLMs) excel in acquiring extensive general knowledge and human-like reasoning capabilities.
- LLM-based agents interact with external tools and environments, playing integral roles in agency through actions, planning, and evaluation.
- The survey aims to clarify the roles of LLMs in agent implementations by creating common workflows that incorporate reusable LMPCs.
- By exploring LLM-based agentic workflows and LMPCs, researchers can effectively leverage these advanced language models for developing intelligent AI agents.

## Summary

Recent advancements in Large Language Models (LLMs) have revolutionized the development of sophisticated agentic workflows. These models have surpassed traditional single-path prompting techniques such as Chain-of-Thought (CoT). This survey delves into the intricacies of these workflows, with a specific emphasis on LLM-Profiled Components (LMPCs), while disregarding non-LLM components. The primary objective is to provide a comprehensive understanding of the roles played by LLMs and assess the reusability of LMPCs. Generative Large Language Models (GLMs or LLMs) have shown remarkable proficiency in acquiring extensive general knowledge and human-like reasoning capabilities. This makes them indispensable in constructing AI agents known as LLM-based agents. These agents actively interact with external tools and environments, functioning as integral components of agency through actions, planning, and evaluation. The motivation for this survey stems from the observation that many LLM-based agents incorporate similar workflows and components despite facing various technical and conceptual challenges such as search algorithms, tree structures, and Reinforcement Learning (RL) components. While existing approaches offer modular solutions or comprehensive reviews of LLM agents' capabilities across profiling, memory, planning, and action; this survey takes a different approach by focusing on the involvement of LLMs within agentic workflows. Instead of aiming to cover all components comprehensively, this survey concentrates on clarifying the roles of LLMs in agent implementations by creating common workflows that incorporate reusable LMPCs. By doing so, it provides a clearer understanding of how LLMs contribute to agentic workflows and highlights their significance in enhancing overall performance. Through this refined exploration of LLM-based agentic workflows and LMPCs, researchers can gain valuable insights into leveraging these advanced language models effectively for developing intelligent AI agents.