Engineering Agentic Network Operations
A hands-on Packet Coders workshop. Engineer an AI agent with the P.E.N.E. framework to build read-only, MCP-exposed network tools to safely troubleshoot networking issues.
Live session + lifetime recording · Full workshop repo · Limited seats
What you'll build
A working AI-assisted troubleshooting workflow
You leave with something real - not a demo. Across the session you'll build, piece by piece:
Ask: "Can you check leaf1 and tell me if anything looks wrong?"
The agent can inspect, summarize and recommend - it works only from evidence collected by the tools you approved. It cannot change the network.
What you'll learn
The most important thing isn't a tool - it's the workflow: prompt → generate → review → run → test → improve → expose.
Generate code with agents
Use AI coding agents to write network automation, with prompts and the P.E.N.E. framework that produce useful, specific code.
Review for risk
Read AI-generated code like a senior engineer - spotting unsafe commands, hardcoded secrets and weak error handling before you run it.
Test & debug
Run generated Python, capture runtime errors and feed them back to the agent to improve the code, test-driven.
Build read-only tools
Add safety constraints and build read-only troubleshooting tools - avoiding dangerous generic command-execution patterns.
Expose tools with MCP
Expose only narrow, approved functions through MCP - and understand why a generic run_command tool is dangerous.
Reason over evidence
Use Claude or OpenAI (BYOK, with a mock-mode fallback) as a reasoning layer over collected evidence - not as a source of truth.
Agenda
Four hours, nine modules
Learn by doing. Hands-on AI network automation, built live.
1 · Welcome & big picture
Chatbots vs scripts vs copilots vs agents - why agents need tools, why the engineer stays critical, and the workshop's safety model.
2 · Lab setup & environment
Repo, Python venv, Docker and an Arista cEOS topology. BYOK API setup with a mock-mode fallback.
3 · Agent context & safety
Write an agent context file: the lab, repo, devices, allowed commands and read-only boundaries - approved vs unsafe tool patterns.
4 · Agent builds the inventory loader
Prompt the agent for a real first component - validating YAML, returning structured data and clear errors. inventory_loader.py
5 · Agent builds read-only network tools
Live device inspection: SSH, read-only show commands, structured JSON.
Break
6 · Review, debug & improve
Review generated code like a senior engineer, run it, capture errors and feed them back. Test-driven improvement and guardrails.
7 · Add the reasoning layer
Send structured network data to Claude/OpenAI to summarize, identify causes and recommend the next safe check - evidence-based, not hallucinated.
8 · Expose approved tools with MCP
Register narrow, approved, read-only functions as MCP tools - and why a generic run_command tool is dangerous.
9 · Final troubleshooting agent demo
Bring it together: the agent uses approved tools, collects evidence and returns a grounded recommendation - without touching config.
Who it's for
You don't need to be an advanced Python developer - just comfortable with basic networking, the CLI, and reading simple code.
What's included
A complete workshop repo to keep
You're not just leaving with a demo - you're leaving with a reusable pattern you can keep learning from after the session.
Your instructor
Sif Baksh
Principal Systems Engineer, Tines
Sif Baksh is a Principal Systems Engineer at Tines with more than 15 years of experience in network automation and cybersecurity. Previously at Swimlane and Infoblox, he helps NetOps and SecOps teams design scalable, resilient automation. Sif is also a technical blogger and volunteer cybersecurity educator.
FAQs
Do I need to know Python?+
You should be able to read basic Python, but you don't need to be an advanced developer. Part of the workshop is learning how to guide an AI coding agent to generate useful Python, then review and test it safely.
Do I need an API key?+
Bring your own Claude or OpenAI API key for the live LLM portions. The workshop supports Claude, OpenAI and mock mode - if you don't have a key, you can still complete the labs using mock mode.
Will the AI make changes to the network?+
No. The workshop uses a read-only safety model. The agent can inspect devices, collect facts, check interface and BGP state, summarize evidence and recommend next checks - it cannot make configuration changes.
Why use MCP?+
MCP gives a clean way to expose approved tools to an agent. Instead of arbitrary commands, you expose specific functions like check_interfaces, check_bgp_neighbors and get_device_facts - safer, easier to test and easier to explain.
Why not just give the agent a generic run command?+
Because that gives the agent too much power and quickly becomes unsafe. This workshop teaches you to expose narrow, approved, read-only tools instead.
Can this work with other vendors?+
Yes. We use Arista cEOS because it's practical for labs, but the same pattern adapts to Cisco, Juniper, Palo Alto, Fortinet, F5, NetBox, Nautobot, Infoblox, Meraki and more: approved tools + structured context + human review + MCP exposure + evidence-based reasoning.
What's the main takeaway?+
AI doesn't replace the network engineer. The engineer becomes the architect, reviewer, tester and safety gate. The agent helps generate code and reason over evidence - but the engineer defines what gets built, what's safe, and what gets exposed.