# Agent-S: LLM Agentic workflow to automate Standard Operating Procedures

**Authors:** Mandar Kulkarni  
**arXiv:** [2503.15520](https://arxiv.org/abs/2503.15520)  
**Date:** 2025-02-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2503.15520v1/

## Key points

- Novel approach to automating Standard Operating Procedures (SOP) using Large Language Models (LLMs)
- Proposed agentic workflow utilizes LLMs augmented with memory and various environments
- Architecture includes task-specific LLMs, Global Action Repository (GAR), execution memory, and multiple environments
- SOP workflow represented as a logical block of text where agent uses current execution memory for next actions
- Fault-tolerance mechanism allows dynamic action repetition or seeking input from external knowledge sources
- Experiments conducted on e-commerce seller domain SOPs to validate agent's performance under complex scenarios
- Successful applications of LLM-based agents in web navigation, travel planning, video understanding, biomedical discovery, code documentation, etc.
- Importance of memory in functionality of LLM-based agents with different methods explored in literature
- Approach offers flexibility in handling SOPs and error management through single-state LLM prompt for predicting actions across all SOPs
- Effectiveness of proposed agentic workflow demonstrated in experimental results for automating SOPs
- Framework's simplicity and adaptability suitable for various applications beyond customer care operations

## Summary

In this paper, we present a novel approach to automating Standard Operating Procedures (SOP) using Large Language Models (LLMs) as the foundation for AI agents. SOPs are crucial in customer care operations as they provide a structured process for human agents to resolve customer issues. Our proposed agentic workflow utilizes LLMs augmented with memory and various environments such as API tools, user interfaces, and external knowledge sources. The architecture of our AI agent includes three task-specific LLMs, a Global Action Repository (GAR), execution memory, and multiple environments. The SOP workflow is represented as a simple logical block of text where the agent uses its current execution memory to determine the next action based on the SOP instructions. It interacts with the appropriate environment (user or API) to gather feedback and observations which are then inputted back into memory to determine subsequent actions. One key feature of our agent is its fault-tolerance mechanism that allows it to dynamically repeat actions or seek input from external knowledge sources when needed. We conducted experiments using three SOPs from the e-commerce seller domain to validate our agent's performance under complex real-world scenarios. Additionally, we discuss how LLM-based agents have been successfully applied in various domains such as web navigation, travel planning, video understanding, biomedical discovery, code documentation, knowledge graph reasoning, and knowledge base question answering. Memory plays a crucial role in these agents' functionality with different methods explored in existing literature. Our approach differs from previous work by offering more flexibility in handling SOPs and error management through a single-state LLM prompt that can predict actions across all SOPs. The experimental results demonstrate the effectiveness of our proposed agentic workflow for automating SOPs and highlight the importance of having an LLM with strong reasoning capabilities. In conclusion, our research presents a novel framework for automating SOP workflows using LLM-based AI agents. The architecture's simplicity and adaptability make it suitable for various applications beyond customer care operations. Further exploration could extend this framework to efficiently automate other types of workflows.