AIMCPNetwork automationVisualizationItentialCommunity

A Network You Can Fly Through: My Conversation on ARQIV

What if you could take the network you operate every day, press Play, and fly through it?

That is where my conversation with Steve Ryan on ARQIV went, after a detour through Battle Chess, generated animations, and the tools we use to build things for fun. By the end, we were discussing how to bring agents into an enterprise, how to pay for them, and what we should teach the next generation of engineers.

ARQIV published He Turned a Real Network Into a 3D World With AI on November 13, 2026. The episode runs 52 minutes. Steve told me near the end that he had prepared hundreds of lines of notes and barely touched them. We kept finding another idea worth following.

Watch the full conversation on YouTube or explore ARQIV and Steve’s other conversations.

From Battle Chess to a network in Unreal Engine

At 1:44, I describe rebuilding Battle Chess with generated pieces and animated encounters. Steve had been using video generation for his own show, so we immediately found common ground: the excitement of making something that previously required an entirely different set of skills, followed by the bill for all those iterations.

That leads to Blender and then, at 4:22, the network example behind the episode title. I explain taking Cisco Discovery Protocol neighbor information, interface details, labels, and IP addresses and representing that topology in Unreal Engine 5 through MCP tooling. The result is a world I can move through, following connections between ports and devices.

The interesting part for me is the relationship between the source data and the view. A topology contains things an engineer already understands. Giving those relationships a spatial representation offers another way to explore them. The conversation also reaches world models and an experimental digital fruit fly trained on a small network-health task. Those are experiments I describe on the show, rather than evidence that a new interface can replace operational testing.

Explore more of my network visualization work.

Start with useful work and limited access

At 12:46, we move from experiments to enterprise operations. My starting point is the same one I would use with a junior engineer: let the agent learn the environment and do useful work with limited permissions.

There is plenty to do before giving an agent permission to change infrastructure. It can read logs and tickets, inspect configurations, compare intended and observed state, and help bring documentation up to date. Policies, procedures, topology, and the available tools all shape whether that work is useful. Read access still needs appropriate controls; it is a starting scope for earning trust.

Steve’s experience with large operational systems makes this discussion concrete. Adding a capable new participant can still disrupt a process if that participant does not understand its dependencies. We return to onboarding, permissions, and the consequences of a mistake throughout the conversation.

Put reasoning around repeatable execution

At 16:27, I describe a pattern I keep returning to at Itential: use an agent to interpret the request and select a workflow, let tested automation perform the change, then use the agent to explain the outcome.

An agent can also help discover how to solve a problem. Once the steps are understood and tested, that work can become a conventional script or workflow. Repeated successful runs help us learn; they do not establish that every edge case has been covered.

This changes the engineer’s work. At 21:41, we discuss managing agents that operate infrastructure. At 28:22, I give an example of a specialized agent that evaluates test results and makes a go/no-go recommendation. Defining its responsibility, the evidence it receives, and who can act on its answer is part of the engineering.

Context and cost are operating concerns

One small idea caught Steve’s attention at 31:44: a canary instruction. Ask the assistant to include a simple, consistent greeting, then watch for that instruction to disappear during a long session. I describe it as a cue to reconsider the session’s context. It is a heuristic, not a test that proves the rest of the work is correct.

At 38:32, we talk about token costs. My approach includes specification-driven development, using local open-weight models for some work, and reserving more expensive models for tasks that benefit from them. We also discuss reducing the size of data sent to models. A specification helps organize the work; it does not automatically make the token bill smaller.

The question we kept coming back to: what should juniors learn?

The most substantial disagreement begins at 41:47. I argue that new engineers need new fundamentals: retrieval, tool calling, MCP, agent coordination, and the ability to build and evaluate these systems. Steve pushes back on the risk of accepting an answer without enough underlying knowledge to judge it.

At 44:41, I hold up the pyATS book and contrast studying a large technical book with connecting a tool through a small JSON configuration. Someone can begin asking useful questions much sooner. Steve’s question remains: who understands the result well enough to recognize when it is wrong?

I do not claim to have resolved that tension. I want more people to participate in network automation, and I want the next generation to have a path into the profession. Steve wants people to understand the risks behind the decisions they approve. The discussion lands on learning how to learn: using the tools available, investigating the problem, and continuing to build judgment as the tools change.

Thank you to Steve for following the ideas, challenging me, and leaving room for that exchange.

Episode date, duration, and chapter links were checked against the YouTube upload. This recap draws on its automatic captions, with technical names normalized for readability. Episode artwork is from ARQIV.

← More field notes