# Beyond State Machines: Executing Network Procedures with Agentic Tool-Calling Sequences

**Authors:** Purna Sai Garigipati, Onur Ayan, Kishor Chandra Joshi, Xueli An  
**arXiv:** [2605.02584](https://arxiv.org/abs/2605.02584)  
**Date:** 2026-05-04  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2605.02584v1/

## Key points

- Study on utilization of Agentic AI in designing future mobile communication systems
- Investigation of four different approaches for agent execution and procedure distribution
- Evaluation using User Equipment (UE) IP allocation procedure as a case study
- Results showed latency and execution correctness differences among the approaches
- Stress test revealed limitations in sequential procedural steps for LLM agents
- Models with advanced tool-calling capabilities demonstrated reliable execution over longer procedures
- Introduction of a procedure-specific error taxonomy to analyze failures in multi-step procedural execution
- Potential of Agentic AI in revolutionizing mobile communication systems by providing flexible services, automating complex network operations, and enabling autonomous decision-making

## Summary

The study conducted by Purna Sai Garigipati, Onur Ayan, Kishor Chandra Joshi, and Xueli An delves into the utilization of Agentic AI in designing future mobile communication systems. The research focuses on how -based network AI agents can execute network procedures through sequences of tool invocations. Four different approaches were investigated to understand how the agent acquires the procedure and distributes execution between itself and underlying tools. Through an evaluation using a User Equipment (UE) IP allocation procedure as a case study, the researchers analyzed latency and execution correctness across the four approaches. A stress test was also conducted to determine the number of sequential procedural steps an LLM agent can reliably execute before failure. The results indicated that approaches relying on iterative agent-side reasoning experienced higher latency and were more susceptible to execution errors. On the other hand, approaches where the procedure was encapsulated within a single tool showed reduced latency by limiting repeated reasoning. Furthermore, the stress-test results revealed that models with advanced tool-calling capabilities maintained reliable execution over longer procedures compared to other evaluated models. However, all models exhibited reliability degradation as procedure length increased, highlighting clear execution limits in multi-step tool-based workflows. To systematically analyze failures in procedure execution, a procedure-specific error taxonomy was introduced to categorize deviations in multi-step procedural execution. In conclusion, this research sheds light on the potential of  in revolutionizing mobile communication systems by providing flexible services, automating complex network operations, and enabling autonomous decision-making. The findings offer valuable insights into optimizing network procedures through  and highlight key considerations for enhancing efficiency and reliability in future mobile communication technologies.