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

Abstract

Recent advancements in Large Language Models (LLMs) have catalyzed the development of sophisticated agentic workflows, offering improvements over traditional single-path, Chain-of-Thought (CoT) prompting techniques. This survey summarize the common workflows, with the particular focus on LLM-Profiled Components (LMPCs) and ignorance of non-LLM components. The reason behind such exploration is to facilitate a clearer understanding of LLM roles and see how reusabile of the LMPCs.

Method

The authors of the survey used a literature review methodology to synthesize and analyze existing research on Large Language Models (LLMs) and their application in intelligent agentic workflows. They examined the roles of LLM-Profiled Components (LMPCs) and how they are integrated into different workflows, such as policy-only, search-based, and feedback-learning workflows. The survey also considered the interaction of LLMs with various environments, including Natural Language Interaction Environments (NLIEs) and tool environments. The authors categorized and detailed the types of workflows and LMPCs, and provided examples of their integration into prominent models. This approach allowed them to identify common patterns and components in LLM-based agents, and to suggest future research directions for improving the autonomy and efficiency of these agents.

Main Finding

The authors discovered that despite the variety of technical and conceptual challenges in the field of LLM-based agents, there are common workflows and components that are widely used. They identified four task-agnostic LMPCs (actors, planners, evaluators, and dynamic models) and other task-dependent LMPCs, which are integral to the functioning of LLM-based agents. The survey also found that these components are organized into three types of modular workflows: policy-only workflows, search-based workflows, and feedback-learning workflows. Additionally, the authors observed that the integration of these components into agentic workflows offers several advantages, such as enhancing understanding, enabling reuse and adaptation, and simplifying modification and extension of existing workflows. They also noted that the design and integration of tools into LLM agents add complexity and require careful consideration of how LLMs interact with both task environments and these auxiliary tools.

Conclusion

The conclusion of the survey is that the understanding of LLM-based agentic workflows and the roles of LLM-Profiled Components (LMPCs) within these workflows is crucial for the development of sophisticated AI agents. The survey summarizes the common workflows and LMPCs, and highlights the importance of these components in facilitating the reuse and adaptation of workflow-level and component-level implementations for constructing complex agents. It also simplifies the modification and extension of existing workflows, as they typically incorporate one or more of these components. The survey encourages further research into the autonomy and efficiency of LLM-based agents, and suggests that future work should focus on reducing the bandwidth required for LLM inference and creating a unified workflow across tasks.

Keywords

Large Language Models, LLM-Profiled Components, LMPCs, intelligent agentic workflows, policy-only workflows, search-based workflows, feedback-learning workflows, Natural Language Interaction Environments, NLIEs, tool environments, actors, planners, evaluators, dynamic models, verbalizers, reusability, adaptability, autonomy, efficiency, methodology, literature review, integration, interaction, environments, tools, components, implementation, future research, bandwidth, unified workflow, memory design, peripheral component integration