As I’ve pointed out in several posts, and as you’ve certainly noticed, there is a teeny bit of hype surrounding AI these days. We’re told network engineers will be obsolete as our AI buddies take over our jobs. Want to roll out a network? No problem, your receptionist will do it for you while sipping a latte, pushing a button to call her AI agent. Some of us take a more realistic view. AI is a tool. An amazing one in many ways, but a tool nonetheless. And the big question is: how useful is it, really?

Most of the folks making the wild claims are marketers, “analysts”, and others who know little to nothing about networking. John Capobianco is certainly not one of these. John knows a lot about networking, and is a truly technical guy. (I’ve never met him in person, so he could be an AI agent, but I hope not.) John recently posted a detailed video about running an AI agent per device, and aggregating these up to a sort of agent-of-agents. His POC demonstrates that his AoA can go out and perform operations on multiple network devices. Cool.

That said, it brings me back to my early days in programmability, 2015 or so. One of the demos we used to run used Alexa. You could give voice commands to Alexa to go configure your routers, or collect data from them. Very cool! How many network engineers to date use Alexa to configure their networks? Approximately zero.

Of course, we weren’t really expecting to kick off a craze of Alexa-network-engineers. The demo was, as my peer Jason Frazier liked to say, about the “art of the possible.” Our model-driven interfaces made it that much easier to connect things in ways that weren’t previously possible.

As I mentioned in a recent post, programmability didn’t always click with network engineers. We would do a demo where we configured VLAN 100 using NETCONF. We just wrapped a giant block of XML in a giant block of Python, and–voilá–VLAN 100! The only problem was, every network engineer who saw the demo rolled his eyes and said, “I could do that with one line of CLI.”

Here’s the question for Capobianco: Is typing “Please configure an NTP server of 192.168.100.100 on all four devices” easier than, say, configuring an NTP server on all four devices? Or even using an Ansible script to do so? Is typing “What is the IP address of Ethernet 0/1 on R1 and R2” better than just going to R1 and R2 and looking at the “sh ip int brief” output?

We must also bear in mind that AI, still, has trouble with more complex tasks. In an earlier post I found it could not configure VRRP for IPv6 correctly. Even after providing it the operational output of the failed config, it still couldn’t provide the correct config. So, NTP servers are fine, but when we get into really complex stuff, will AI shine or fail?

I’ve been finding ChatGPT incredibly helpful working in my lab lately. I needed to do a few things on ProxMox, and it saved me a lot of searching through StackExchange. But if we want to go from useful tool to fully agentic network management, well, we have a long way to go. Right now the agentic demos feel a bit like the Alexa ones–the art of the possible, but not necessarily the probable.