Telco RCA AI Agent

This repository contains the work for a telecom root-cause analysis (RCA) agent challenge spanning two tracks:

• Wireless_Troubleshooting focuses on wireless-network troubleshooting and optimization.
• IP_Network_Troubleshooting focuses on IP-network troubleshooting and fault localization.

The goal is to build agentic systems that can reason over structured telecom data, interact with tool/simulation servers, and produce accurate troubleshooting answers under competition constraints.

Project scope

This repo includes:

• agent implementations and supporting code for Wireless_Troubleshooting and IP_Network_Troubleshooting
• local server and data utilities
• evaluation and submission scripts
• architecture notes and design summaries for both agent approaches
• example outputs and competition artifacts

Repository layout

• Wireless_Troubleshooting — wireless troubleshooting agent code, utilities, and track-specific documentation
• IP_Network_Troubleshooting — IP-network troubleshooting agent code, CLI/server logic, and evaluation tools
• architecture — design summaries and architecture notes for the agent strategies
• submission — example and generated submission files
• sources — reference material and background notes

Wireless_Troubleshooting agent summary

Wireless_Troubleshooting uses a structured, evidence-first controller for wireless RCA. The workflow is designed to be mostly deterministic:

1. Load the scenario and parse task metadata.
2. Retrieve telemetry from local data or the provided server APIs.
3. Extract wireless evidence such as poor-throughput periods, serving cells, neighbor relationships, and KPI anomalies.
4. Parse and score candidate actions.
5. Use the language model only for targeted adjudication when ambiguity remains.

This keeps the agent reliable, faster, and easier to audit than a fully free-form reasoning loop.

IP_Network_Troubleshooting agent summary

IP_Network_Troubleshooting uses a hybrid architecture that combines deterministic network reasoning with bounded model assistance. The system is organized around:

1. task classification and entity extraction
2. evidence planning and CLI command execution
3. topology and forwarding-graph construction
4. rule-based or graph-based solving for path and fault-localization questions
5. model-assisted repair or final adjudication only when needed

This architecture is intended to balance precision, traceability, and efficiency for multi-vendor IP-network troubleshooting.

Architecture references

For the detailed design notes behind each agent approach, see:

• architecture/track1_mainpy_architecture.md — Wireless_Troubleshooting architecture summary
• architecture/trackb_architecture_option_4_hybrid_competition_agent.md — IP_Network_Troubleshooting hybrid architecture summary
• architecture/rcagent_concrete_architecture.md — a concrete unified agent design proposal

Notes

• The competition uses Qwen3.5-35B-A3B as the base model.
• The repository is organized around two separate tracks with different task formats and tooling.
• For implementation-specific details, see the track-specific READMEs and the architecture notes in architecture.

Getting started

• Review Wireless_Troubleshooting/README.md for the wireless track.
• Review IP_Network_Troubleshooting/README.md for the IP-network track.
• Review the architecture summaries in architecture for the underlying design choices.