Source-linked AI summary
ProAgent: From Robotic Process Automation to Agentic Process Automation
Yining Ye, Xin Cong, Shizuo Tian, Jiannan Cao, Hao Wang, Yujia Qin, Yaxi Lu, Heyang Yu, Huadong Wang, Yankai Lin, Zhiyuan Liu, Maosong Sun
TL;DR
RPA still depends on people to design elaborate workflows and handle dynamic decisions that rule-based execution cannot express. The paper proposes Agentic Process Automation and instantiates ProAgent, an LLM-based agent framework for workflow construction and execution; proof-of-concept experiments provide evidence of feasibility. The paper also notes risks of over-reliance on agent decisions, including automation bias.
Problem
RPA requires human intelligence for elaborate workflow design and struggles with complex processes that need dynamic decisions during execution.
Source: §1 · p. 2Method
The paper proposes Agentic Process Automation and instantiates ProAgent, which uses LLM-based agents to construct workflows from instructions and handle dynamic decisions during execution.
Source: §1 · p. 2Results
Proof-of-concept experiments provide evidence that Agentic Process Automation can feasibly support process automation through workflow construction and complex decision-making by agents.
Source: §6 · p. 11Takeaways & Limitations
The authors conclude that their proof-of-concept provides evidence of feasibility in achieving efficiency and flexibility in process automation.
Source: §6 · p. 11Agent-based automation may intensify automation bias, as people may trust agent decisions even when they conflict with their views or involve hallucinations.
Source: §4.3 · p. 9
Abstract
from arXiv · showhide
From ancient water wheels to robotic process automation (RPA), automation technology has evolved throughout history to liberate human beings from arduous tasks. Yet, RPA struggles with tasks needing human-like intelligence, especially in elaborate design of workflow construction and dynamic decision-making in workflow execution. As Large Language Models (LLMs) have emerged human-like intelligence, this paper introduces Agentic Process Automation (APA), a groundbreaking automation paradigm using LLM-based agents for advanced automation by offloading the human labor to agents associated with construction and execution. We then instantiate ProAgent, an LLM-based agent designed to craft workflows from human instructions and make intricate decisions by coordinating specialized agents. Empirical experiments are conducted to detail its construction and execution procedure of workflow, showcasing the feasibility of APA, unveiling the possibility of a new paradigm of automation driven by agents. Our code is public at https://github.com/OpenBMB/ProAgent.
1 INTRODUCTION
RPA automates rule-based workflows but still requires human intelligence to design elaborate workflows and handle dynamic decisions. The paper proposes APA and instantiates ProAgent to shift workflow construction and dynamic execution decisions to LLM-based agents.
1 INTRODUCTION
- RPA workflows still require human intelligence for elaborate design and struggle with complex tasks needing dynamic decisions during execution.RPA automates processes by orchestrating software through manually crafted rules into solidified workflows.Source: §1 · p. 2
- APA assigns LLM-based agents to construct workflows from human instructions and intervene in dynamic decision-making during execution.Agents identify where dynamic decisions are needed and are orchestrated into the workflow.Source: §1 · p. 2
- ProAgent uses an Agentic Workflow Description Language and DataAgent and ControlAgent to construct workflows and handle complex data processing and branching.The language uses JSON for input/output data and Python for process control logic.Source: §1 · pp. 2–3
- Proof-of-concept experiments demonstrate ProAgent constructing workflows from human instructions and handling dynamic decisions with agents in the workflow.The paper also explores APA’s potential across tool learning, process mining, and safety.Source: §1 · p. 3
2 METHODOLOGY
The methodology represents workflows in a form intended for LLM agents, integrates agents into data and control flow, and describes construction and execution procedures. ProAgent uses JSON and Python, specialized agents, and iterative workflow-generation operations.
2 METHODOLOGY
- The methodology presents a workflow representation, agent integration, and procedures for workflow construction and execution.Workflows solidify processes through software-operation nodes and execution-topology edges, with data flow and control flow.Source: §2 · p. 3
2.1 AGENTIC WORKFLOW DESCRIPTION LANGUAGE
- Agentic Workflow Description Language uses JSON to standardize action inputs and outputs and Python to express control logic such as branches, loops, and nested workflows.Actions map to Python functions, with workflow execution starting at mainWorkflow.Source: §2.1 · pp. 3–4
2.2 AGENT-INTEGRATED WORKFLOW
- DataAgent handles intricate data processing, while ControlAgent uses a natural-language judgment criterion to select a subsequent execution branch.DataAgent’s inputs and outputs remain organized as JSON so it can connect with other workflow actions.Source: §2.2 · p. 5
2.3 WORKFLOW CONSTRUCTION
- Workflow construction iterates through action definition, action implementation, workflow implementation, and task submission.The described implementation uses GPT-4, tests functions or workflows during construction, and employs function calling and plans with comments.Source: §2.3 · pp. 5–6
2.4 WORKFLOW EXECUTION
- Workflow execution uses a Python interpreter to run mainWorkflow from its entry point according to Python’s sequential code-execution order.Execution finishes successfully when mainWorkflow returns.Source: §2.4 · p. 6
3 PROOF-OF-CONCEPT EXPERIMENT
A proof-of-concept commercial scenario tests ProAgent’s construction and execution of a workflow for classifying business lines and routing their reports. The case illustrates a seven-node workflow and agent decisions for two business-line types.
3.1 TASK CONSTRUCTION
- The commercial task reads business-line data from Google Sheets and sends To-Customer profits to Slack or a To-Business analysis and profitability overview to its manager.The case uses n8n actions and an implementation based on GPT-4.Source: §3.1 · p. 6, §3 · p. 6
- The task is reusable across business departments with similar spreadsheet formats, while business-line classification and corresponding email generation require agent-based decisions and processing.The paper describes classification as difficult to handle through rule-based judgments.Source: §3.1 · p. 6
3.2 WORKFLOW CONSTRUCTION
- The case workflow has seven nodes, including DataAgent and ControlAgent, and uses a loop to process multiple To-Business data entries.Its actions include a webhook trigger, Google Sheets reading, conditional routing, and branch-specific actions.Source: §3.2 · pp. 7–8
3.3 WORKFLOW EXECUTION
- In two execution cases, ControlAgent distinguishes To-Business from To-Customer, triggering email generation for the former and a Slack message for the latter.For the To-Business ERP example, DataAgent generates an email reporting revenue and positive feedback.Source: §3.3 · pp. 8–9
4 DISCUSSION
The discussion connects ProAgent to tool use and workflow improvement, then considers automation bias and the continuing role of human judgment.
4.1 PROAGENT AS TOOLAGENT: INTEGRATING TOOL UTILIZATION AND TOOL CREATION
- ProAgent combines workflow construction, understood as tool creation, with workflow execution, understood as tool utilization.The authors frame it as an autonomous process integrating both.Source: §4.1 · p. 9
4.2 PROCESS MINING: DISCOVER, ANALYZE, AND IMPROVE WORKFLOWS
- Agent-only workflow construction may yield sub-optimal workflows; Process Mining can use execution histories to identify inefficiencies and inform optimization.The optimized workflows may also provide training data for improving agent construction.Source: §4.2 · p. 9
4.3 ETHICAL AND SAFETY CONCERNS: AGENT AGGRAVATES AUTOMATION BIAS
- Automation Bias arises when people trust agent decisions despite conflicts with their views, and LLM hallucinations can make that over-reliance more pronounced.The authors call for safer, more trustworthy, and more interpretable agentic process automation.Source: §4.3 · pp. 9–10
4.4 HUMAN ADVANTAGE: RETHINKING THE MEANING OF HUMAN LABOR
- Processes involving intuition, experience, and creativity may benefit from human involvement, so APA’s frontier includes deciding where oversight or intervention remains appropriate.The authors frame automation as amplifying productivity rather than wholly supplanting humans.Source: §4.4 · p. 10
5 RELATED WORK
The related work contrasts rule-based RPA and earlier AI additions with APA’s use of LLM-based agents in workflow construction and execution. It also positions ProAgent against prior agent approaches that generate decisions linearly.
Robotic Process Automation
- RPA automates repetitive, rule-based tasks with predefined workflows, but manual rules make complex dynamic decision-making difficult.RPA can automate software APIs or simulate GUI interactions without transforming existing software assets.Source: §5 · p. 10
AI-integrated process automation
- IPA and Cognitive Automation add AI to specific RPA phases, whereas APA integrates LLMs into autonomous workflow generation and agent-driven execution.Examples of targeted phases include data transformation, workflow optimization, conversational assistance, and demonstration-to-process translation.Source: §5 · p. 10
LLM-based Agents
- Early LLM-agent approaches generate actions from observations, while ReAct (Yao et al., 2022b) adds reasoning chains; later work explores reflection, planning, tool learning, and multi-agent methods.The passage characterizes ReAct’s reasoning-guided actions as flexible, trackable, and adjustable compared with act-only methods.Source: §5 · pp. 10–11
- Existing ReAct-based methods are described as linearly generating decisions; ProAgent instead identifies workflow parts needing intelligence and integrates agents there purposefully.The authors present this as an approach to improving dynamic task-solving efficiency.Source: §5 · p. 11
6 CONCLUSION
The conclusion presents APA as an LLM-agent-based process automation paradigm, instantiated by ProAgent. Proof-of-concept experiments support its feasibility for complex decision-making and flexible process automation.
6 CONCLUSION
- APA integrates LLM-based agents into workflow construction and execution to address RPA’s limitations on tasks requiring human intelligence.The authors frame this as a way to offload intelligent labor from humans.Source: §6 · p. 11
- ProAgent illustrates the feasibility of using LLM-based agents to manage complex decision-making processes.The paper reports proof-of-concept experiments as evidence for that feasibility.Source: §6 · p. 11
- The conclusion identifies efficiency and flexibility in process automation as outcomes supported by the proof-of-concept experiment.It anticipates APA may support further automation advances, including reduced human intervention.Source: §6 · p. 11