Uploaded June 2026 | Updated September 2026, 1 week ago
Abstract
Everyone's talking about AI agents. Few are building them for real network operations.
This talk walks through the practical journey of building network agents. From "should I even do this?" to production tradeoffs. We'll answer three questions network engineers are actually asking:
1. Do I need to build an AI Agent?
Tools like Claude Code and ChatGPT can already call APIs and parse output. When is that enough? When does it break down? I'll show where off-the-shelf tools hit their limits with real network scenarios.
2. If I build an agent should it be hosted or self-hosted?
Sending packet captures to OpenAI isn't an option for most of us. But self-hosted models come with their own challenges. I'll demo the same network troubleshooting workflow with Claude Code and Llama running locally to show exactly where each succeeds and fails.
3. What are the actual tradeoffs?
Cost, latency, accuracy, data privacy, and the one nobody talks about: how much of your time it takes to build and maintain. I'll share what I learned building my own network agent with a network-mcp, an open-source MCP server that gives AI agents network diagnostic tools.
You'll leave with a framework for deciding what to build, what to buy, and what to skip, based on your actual operational needs, not hype.
nanog.org/events/nanog-97/content/5687
Abstract
Everyone's talking about AI agents. Few are building them for real network operations.
This talk walks through the practical journey of building network agents. From "should I even do this?" to production tradeoffs. We'll answer three questions network engineers are actually asking:
1. Do I need to build an AI Agent?
Tools like Claude Code and ChatGPT can already call APIs and parse output. When is that enough? When does it break down? I'll show where off-the-shelf tools hit their limits with real network scenarios.
2. If I build an agent should it be hosted or self-hosted?
Sending packet captures to OpenAI isn't an option for most of us. But self-hosted models come with their own challenges. I'll demo the same network troubleshooting workflow with Claude Code and Llama running locally to show exactly where each succeeds and fails.
3. What are the actual tradeoffs?
Cost, latency, accuracy, data privacy, and the one nobody talks about: how much of your time it takes to build and maintain. I'll share what I learned building my own network agent with a network-mcp, an open-source MCP server that gives AI agents network diagnostic tools.
You'll leave with a framework for deciding what to build, what to buy, and what to skip, based on your actual operational needs, not hype.
nanog.org/events/nanog-97/content/5687
