ARTFEED — Contemporary Art Intelligence

SADE: LLM Agent for Network Troubleshooting Using Cisco Methodology

ai-technology · 2026-05-07

A novel agent utilizing LLM technology, named SADE (Symptom-Aware Diagnostic Escalation), has been created to enhance root-cause identification in network troubleshooting. This agent incorporates the traditional Cisco troubleshooting approach as a defined policy, implementing a phase-gated diagnostic process that distinguishes between evidence gathering and hypothesis formulation. Additionally, it features a routed library of fault-family skills alongside effective diagnostic aids. In testing on a set of 523 incidents from the public NIKA benchmark, which included eleven previously unseen scenarios, SADE achieved a 37 percentage point improvement in root-cause F1 compared to a ReAct + GPT-5 baseline. The paper contends that current LLM agents are inadequate due to their reliance on unstructured deliberation, unlike the systematic methods employed by human network engineers.

Key facts

  • SADE stands for Symptom-Aware Diagnostic Escalation
  • It encodes the classical Cisco troubleshooting methodology as an explicit policy
  • Uses a phase-gated diagnostic workflow separating evidence acquisition from hypothesis commitment
  • Includes a routed library of fault-family skills and high-yield diagnostic helpers
  • Tested on a held-out 523 incident set from the NIKA benchmark
  • Covers eleven unseen scenarios
  • Improves root-cause F1 by 37 percentage points over ReAct + GPT-5 baseline
  • Paper published on arXiv with ID 2605.04530

Entities

Institutions

  • arXiv

Sources