Back to Blog
AI AgentsNetwork AnalysisTelecomHarness

AI Agents for Telecom

July 20, 20265 min read
0:00 / 0:00

The Status Quo

The client wanted to create AI-powered tools to help junior engineers diagnose network problems faster. The off-the-shelf tool being used, Wireshark — is a domain-specific tool that takes years of expertise to master.

When a critical service is degrading, that's time the business doesn't have.

The Goal

Build an AI-powered system that could automate network packet analysis

Why was the API slow between 2 PM and 2:10 PM?
Which devices experienced the most packet loss?
Is this a network problem or an app problem?
Show me the timeline of failures.

The system should return a structured answer — in plain language — with enough detail for an expert to validate the findings.

Challenge: LLMs Can't Read Network Traffic

Network traffic is a special kind of data format which CANNOT be directly read by LLMs like Claude and ChatGPT. Unlike other data types like text files or PDFs, this is quite niche — and there are no mainstream AI clients which support it out of the box.

The bottleneck isn't the model — it's the context.

With the right harness, an LLM can convert raw binary data from a network stream in a parseable format which it can then actually reason about.

How the Harness Works

We built a three-layer pipeline that turns a raw data into queryable intelligence:

1
1. Data Ingestion + Parsing

We used TShark to parse every packets from LIVE network flows. Our harness allows AI agents to access this data via a custom MCP server

2
2. Proprietary Business Logic

We integrated the Teleco's EXISTING Network rules into the Agent. Now AI could securely utlizes proprietary know-how without Human Intervention.

3
3. Customised UI Dashboard

We built a custom dashbord which tracked Raw Traffic flows with Agentic insights. Engineers could directly highlight traffic flows and ASK question in Natural Lanugage.

The harness acts as a translation layer.

It doesn't make the model smarter — it makes the data accessible. It extracts the signal, filters the noise, and structures the insights into a format the AI can reason about. This is the same principle that applies to every domain we work in at harnes.io:

structured context is the unlock.

The Impact: Before vs. After

Before the Harness

  • Capture PCAP, open in Wireshark
  • Write complex display filters
  • Manually trace TCP streams
  • Cross-reference protocol RFCs
  • Piece together a timeline by hand
  • Hours to find root cause

After the Harness

  • Upload PCAP to the harness
  • Ask a question in plain English
  • AI correlates signals across layers
  • Get structured RCA + visualizations
  • Validate with exact IPs and timestamps
  • Minutes to find root cause

The Vision

Network troubleshooting shouldn't require hours of manual filtering. A support engineer should be able to ask "What went wrong?" and get a clear answer — with root cause, evidence, and a path to validation. The harness handles the grunt work so engineers can focus on judgment and action.

That's the power of a well-designed harness. And it's exactly what we built for our telecom client.

deploy@harnes.io