LLM-Agent-UMF：基于LLM的代理统一建模框架，实现多主动/被动核心代理的无缝集成

LLM-Agent-UMF: LLM-based Agent Unified Modeling Framework for Seamless Integration of Multi Active/Passive Core-Agents

September 17, 2024

作者: Amine B. Hassouna, Hana Chaari, Ines Belhaj

cs.AI

摘要

基于LLM的代理程序中工具的集成克服了独立LLM和传统代理程序功能有限的困难。然而，这些技术的结合以及在几项最新工作中提出的增强方案遵循了非统一的软件架构，导致缺乏模块化。事实上，它们主要关注功能，而忽视了代理程序内部组件边界的定义。这导致研究人员之间术语和架构上的不确定性，我们通过提出一个统一框架来解决这些问题。该框架从功能和软件架构的角度为基于LLM的代理程序的开发建立了清晰的基础。 我们的框架，LLM-Agent-UMF（基于LLM的代理程序统一建模框架），明确区分了代理程序的不同组件，将LLM和工具与新引入的核心代理元素区分开来，核心代理扮演着代理程序的中央协调员的角色，包括五个模块：规划、记忆、配置文件、动作和安全性，后者在先前的工作中经常被忽视。核心代理的内部结构差异导致我们将其分类为被动和主动类型。基于此，我们提出了不同的多核代理架构，结合了各种个体代理的独特特征。 为了评估目的，我们将该框架应用于一些最新代理程序，从而展示其与它们的功能的一致性，并澄清被忽视的架构方面。此外，我们通过将不同代理整合到混合主动/被动核心代理系统中，对我们提出的四种架构进行了彻底评估。这种分析为潜在改进提供了明确见解，并突出了结合特定代理程序所涉及的挑战。

English

The integration of tools in LLM-based agents overcame the difficulties of standalone LLMs and traditional agents' limited capabilities. However, the conjunction of these technologies and the proposed enhancements in several state-of-the-art works followed a non-unified software architecture resulting in a lack of modularity. Indeed, they focused mainly on functionalities and overlooked the definition of the component's boundaries within the agent. This caused terminological and architectural ambiguities between researchers which we addressed in this paper by proposing a unified framework that establishes a clear foundation for LLM-based agents' development from both functional and software architectural perspectives. Our framework, LLM-Agent-UMF (LLM-based Agent Unified Modeling Framework), clearly distinguishes between the different components of an agent, setting LLMs, and tools apart from a newly introduced element: the core-agent, playing the role of the central coordinator of the agent which comprises five modules: planning, memory, profile, action, and security, the latter often neglected in previous works. Differences in the internal structure of core-agents led us to classify them into a taxonomy of passive and active types. Based on this, we proposed different multi-core agent architectures combining unique characteristics of various individual agents. For evaluation purposes, we applied this framework to a selection of state-of-the-art agents, thereby demonstrating its alignment with their functionalities and clarifying the overlooked architectural aspects. Moreover, we thoroughly assessed four of our proposed architectures by integrating distinctive agents into hybrid active/passive core-agents' systems. This analysis provided clear insights into potential improvements and highlighted the challenges involved in the combination of specific agents.