LIVE · 5-day hands-on cohort

Agentic AI for
Network Automation Bootcamp

Five days on a live lab network, from your first LLM API call to a multi-agent change pipeline. Every change starts as a structured, validated proposal, gated for approval, deployed, then independently verified.

Rick Donato
Your instructor Rick Donato · Founder, Packet Coders
Dates 28–30 Oct · 10–11 Dec 2026
Time 13:00–21:00 GMT · 08:00–16:00 EST
Length 5 days
Seats 20 seats

Cloud lab per attendee · Bring your own API key · Take-home repo

CHANGE REQUEST > add vlan 250 to   leaf1 and leaf2 MONITORING ALERT { "alert":   "bgp_down" } AGENT intent AGENT incident AGENT change gate fabric AGENT assurance

Getting a model to write config is the easy part. The engineering problem starts one step later: how does a change earn its way onto the network?

You already gate changes with peer review, a source of truth and pre/post checks. This bootcamp doesn't replace any of that - it drops an agent into the pipeline you already trust.

01 Intent or an alert goes in one end
02 Out comes a structured proposal, materialised as a NetBox branch
03 Diffed against your source of truth, then gated for approval
04 Deployed, then verified by a second agent that didn't write it

Nothing reaches a device that wouldn't survive your change review - and you build every stage yourself.

Outcomes

What you'll walk out able to do

Five days, hands on keys throughout. By the end of the week you can:

Call LLM APIs with structured outputs and tool calling - and control what it costs
Audit configs against your own standards and get typed violations back, with remediation attached
Build and secure MCP servers over live devices and your source of truth
Build RAG pipelines that answer from your own runbooks, docs and configs - with citations, not guesses
Query a topology knowledge graph for answers vector search can't reach
Drive Claude Code as an off-the-shelf NetOps agent - runbooks as skills, guardrail hooks, headless runs
Build agent graphs with approval nodes, checkpointing and commit-confirmed rollback
Wire agents into a multi-agent system over A2A, with separation of duties between them
Turn a plain-English change request into a validated, gated, verified deployment
Take a workflow from human-approved to fully automatic - and justify the controls that make it safe
Test and evaluate AI systems like software - prompt regression tests in CI
Trace, cost-track and observe agent runs in production

The architecture you'll build

Two triggers, one pipeline

A planned change and a 3AM incident look like different jobs. Only the trigger differs. Both end up as the same structured proposal, and from there every change takes an identical path to the network - so you build the controls once and both triggers inherit them.

Trigger 01

Intent · planned

"Add VLAN 250 to leaf1 and leaf2, tag it on the uplinks." An engineer asks in plain English; the agent returns a structured proposal.

Trigger 02

Incident · unplanned

Your monitoring stack POSTs an alert to a webhook. The agent reads the payload and works out what happened - no parser to write per alert source.

Curriculum

Five days. Nothing you build gets thrown away.

Each lab is a component of the final pipeline, not an exercise you bin - by the last day they're wired together.

Module 1

LLM foundations & cost engineering

Typed output, tool calls, and a bill you can predict.

Prompting as engineering - templates in code, versioning, chain-of-thought vs reasoning models
Structured outputs two ways: raw JSON Schema on the wire, then Pydantic contracts
Tool calling against the raw API - schemas, the call loop, and handling bad arguments
Hybrid parsing - TextFSM first, LLM for the messy remainder
Model routing, token counting, prompt caching, batch APIs, rate-limit handling

Lab · Config Intelligence Tool - typed configs in, misconfigurations flagged, remediation generated, cost report per run.

Module 2

MCP - standardised tool access

Read your network conversationally, without handing over write access.

MCP architecture - servers, tools, resources, transports, and how to debug them
Building servers with FastMCP - typed tools, docstrings-as-prompts
A device MCP server over Scrapli/gNMI - state, config, diagnostics
A NetBox server behind a thin SoT-agnostic adapter - intended vs actual state
Security - read-only by default, write allowlists, audit logging, prompt injection via device output

Lab · NetOps MCP Suite - two servers wired into Claude Code, troubleshooting a spine-leaf fabric conversationally.

Module 3

RAG & knowledge graphs - local context retrieval

Grounded answers, with the sources attached.

Embeddings and chunking strategies - where most RAG quietly fails
A full RAG pipeline with LlamaIndex and a vector DB - grounding and citation
Graph RAG over topology - sync your source of truth into a graph DB, then traverse: "what breaks if leaf1 dies?"
The other GraphRAG - why LLM-extracted entity graphs are the wrong tool when a source of truth already holds the topology precisely
Agentic RAG - retrieval as a tool, with the agent choosing its source per question

Lab · NetOps Knowledge Assistant - routes across runbooks, the topology graph and live device state. Ask it what breaks if a device dies, and it traverses the topology to tell you.

Module 4

Agents & agent graphs - reason, act, verify

Closing the loop safely. The heaviest, best day of the week.

Claude Code as a harness - skills, guardrail hooks, read-only subagents, headless runs
Then build: the agent loop, and Pydantic AI in anger - typed results, retries, streaming
Agent graphs in LangGraph - explicit state, branches, approval nodes, checkpointing
Verification at both ends - pre-flight against the proposed change, post-flight against the network
Assertions as pytest against the live fabric, commit-confirmed rollback, and the same checks on a schedule = drift detection
Testing AI systems - pytest for prompts, eval sets from your own scenarios, gating in CI

Lab · Incident Agent - solve a BGP incident with the harness first, then build the agent graph that emits a typed proposal into a NetBox branch. It never touches a device.

Module 5

Multi-agent systems, A2A & capstone

From one agent to a system serving both planned and unplanned change.

Multi-agent patterns - orchestrator/worker, handoffs, and when one agent is enough
Two triggers, one pipeline, and where open loop becomes closed loop
A2A - agent cards, tasks, artifacts, and how it differs from MCP: peers vs tools
Build the Intent Agent, and a webhook receiver thin enough that the agent does the interpreting
Production - tracing, per-run cost tracking, versioning, failure modes, on-prem options

Capstone - the multi-agent NetOps pipeline, built and demoed live.

Tools you'll learn

One opinionated choice per layer, always taught after the raw concepts - so you can swap any of them later.

Model layer

LLM APIs JSON Schema Pydantic

Tool access

MCP FastMCP

Retrieval

LlamaIndex Vector DB Graph DB

Harnesses

Claude Code Skills Hooks

Agents

Pydantic AI LangGraph A2A

Why this, not a course you watch

The parts a demo skips

Multi-vendor devices that break, approval gates, independent verification, rollback and cost control. The unglamorous half is the half that decides whether any of this runs in production.

A live lab network

Your own cloud VM with a multi-vendor fabric, source of truth, vector store and graph database - pre-seeded and ready.

Language in, deployed change out

A pipeline that takes a plain-English request or an alert and ends in a validated, gated, verified change - or an automatic rollback.

Verification at both ends

Pre-flight before anything ships, post-flight after it does - with the verdict enforced by the source of truth, not by the agent that might bypass it.

Cost engineering as a habit

Model routing tables, token counting, caching and batch APIs. No batch operation without a cost estimate first.

A live cohort, capped at 20

Small enough that you get unstuck in minutes. Every lab is scaffolded so you write the interesting parts, not the boilerplate.

A repo you keep working in

Labs, reference solutions, shared helpers, eval harness, lab environment and slides. The starting point for your own internal tooling.

Who it's for

Network engineers who are comfortable writing basic Python and want to build AI-driven network operations tooling - not just chat with an LLM.

Network engineers who've automated with scripts and want the next layer.
Automation engineers asked to "do something with AI" and wanting it done properly.
NetDevOps teams building internal platforms and agent tooling.
Engineers who've used someone else's MCP server and want to build and secure their own.
Graduates of the Python Network Automation Bootcamp looking for what's next.
Technical leaders deciding where agents fit in their NetOps stack.

Not for you if you've never written Python, or if you want a vendor product demo rather than to build it yourself.

Rick Donato

Your instructor

Rick Donato

Founder, Packet Coders

Rick has spent over twenty years in networking and network automation, and founded Packet Coders to deliver hands-on training by engineers, for engineers. He has taught Python network automation to hundreds of engineers through the Packet Coders bootcamps and tech sessions.

This bootcamp is the material he wanted when agents first landed: opinionated framework choices, real devices, and a safety model you can defend in a change review.

FAQs

How much Python do I need?+

You should be comfortable writing basic Python - functions, dictionaries, working with an API, reading a traceback. You don't need to be a software engineer.

Do I need my own lab or API key?+

The lab and development environment are fully provided - a multi-vendor fabric, source of truth, vector store and graph database, all pre-installed and pre-seeded. All you need is a Chrome browser. You bring your own LLM API key.

Will the agents make changes to a network?+

Yes - in your isolated lab, and never directly. Agents emit a typed proposal into a branch of the source of truth. It's validated before deployment, gated on a human merge, deployed with commit-confirmed so the device reverts itself if nothing confirms, then verified independently.

Is this vendor-specific?+

No - you'll perform changes against a multi-vendor lab environment. NetBox is the source of truth, sitting behind a thin adapter interface, so the patterns move to other sources of truth without rewriting the agents.

Which frameworks do you teach?+

One opinionated choice per layer, always after the raw concepts: Pydantic for structured output, FastMCP for tool servers, LlamaIndex for RAG, a graph database for topology, Pydantic AI for agents, LangGraph for agent graphs, A2A for agent-to-agent.

Can my employer pay, and do you do team rates?+

Yes. Invoices for teams are no problem, and teams booking together get a discount on future sessions. Get in touch and we'll sort it.

Reserve your seat

Five days live, your own lab network, and a multi-agent change pipeline you built and can defend in a change review.

Five days live, hands-on, capped at 20 seats
Cloud lab and dev environment provided - bring your own API key
Full repo: five labs, reference solutions, eval harness
Lifetime access to the recordings
Early Bird · 30% Off · Ends 14th Oct
£1746 £2,495

One-time

CodeAIBOOTCAMP30

Apply at checkout

Modules 1–3 · 28–30 Oct 2026

Modules 4–5 · 10–11 Dec 2026

13:00–21:00 GMT · 08:00–16:00 EST · 14:00–22:00 CET