HN837: Agentic AI to Reduce MTTR

Ethan Banks (0:00)

If you're managing a messy multi-vendor network that grew for years and there was no plan, you are fighting unpredictability. When something breaks, METER delivers a complete predictable network, hardware, software and services that are consumed as a subscription. The transition is easy. They'll give you upgrade credits, you get a fully managed install and deployment plus 24-7 support, companies such as Lyft and Bridgewater. They're METER shops.

Go to meter.com/heavynetworking.

That's meter.com/heavynetworking to book a demo. Welcome to Heavy Networking. I'm Ethan Banks, and I find myself in Munich, Germany for the Autocon 5 conference put on by the Network Automation Forum. And if you're interested in that show, visit networkautomation.forum to find out about the next Autocon event coming up in the fall of 2026 That's going to be held in North America. And hey, here I am at Autocon 5 with a lot of folks who love network automation. One of those folks is Eduard Dulharu. Eduard has been working with agents powered by artificial intelligence to improve mean time to recovery, MTTR. And since agentic AI has been talked about in hand-wavy terms by a lot of people, Eduard wants to discuss it in a very practical hands-on way. And so I thought this would be a great conversation for us to have. Eduard, you did a workshop here yesterday at Autocon 5 and that went well, I hear.

Eduard Dulharu (1:24)

Yes, it was a very, very interesting experience to see how many people are fascinated about the power of agentic AI and with real applications for for networking engineers, especially.

Ethan Banks (1:39)

Fascinated and scared. Is agentic AI going to take our jobs, Eduard?

Eduard Dulharu (1:43)

No, I think will, agentic AI will promote engineers from firefighters to system architects, which is my vision. I don't believe a second that agentic AI should replace anybody, but should empower them and should give engineers extra power. Okay.

Ethan Banks (2:05)

Well, and that's kind of the way I've begun thinking about it. Agentic AI is another tool in your toolbox. It's another way to make yourself more efficient.

Still, there's some intimidation there. But your workshop was about agentic AI to reduce MTTR. Can you talk about, let's assume that the people listening know what agentic AI is because we've had done several shows on it where we've talked about agents, and MCP servers, and so on. If we assume we know roughly what agentic AI is, what does agentic AI for MTTR reduction specifically mean?

Eduard Dulharu (2:44)

For MTTR specifically, agentic AI represents a system which reproduces the way in which people do their work and operate.

And then the idea is that all the incidents, especially in complex environments, like I showed yesterday, with multi-vendor fabric, with XANI VPN, with Cisco and Arista, agentic AI can reduce the MTTR from hours in complex troubleshooting scenarios to seconds. So to answer more directly to your question, agentic AI represents a well-deterministic workflow in which the syslogs or telemetry data coming from the environment is analyzed in almost real time by agents using what we called a single source of truth, a catalog with issues. And then based on MCP, we collect real time data to correlate with that specific incident. And all of this information represents the context for a fine-tuned local model which does reasoning across all of this data.

And this reasoning model proposes back to the agent a recommendation, a solution, which then is tested in a digital twin, a digital twin which is running on ContainerLab, which is a one-to-one copy of the environment. And the proposed change is assessed, is tested against the digital twin. And if the confidence level of the test or if the test is successful, it is reported to the engineer in a dashboard. And the engineer receives the problem in real time, the evidence coming from MCP, the reasoning trace from the model, and the proposed solution and the evidence from the digital twin that it was successful. This is the whole system.

Ethan Banks (4:47)

Okay, so when we talk about MTTR, what I'm usually thinking about is recovery from a problem. So I've heard a bunch of presentations in network automation on the past. I've been to all the Autocon conferences. And this particular kind of a presentation will go along the lines of, this is how we reduce MTTR by dispatching a bunch of scripts that are going to do system data gathering for me. Python scripts are going to run, or we're going to pull some show run, or show whatever kind of material from the Cisco box, in this case, or whatever. Parse it and then make a report that's going to tell us a bunch of things. And then that's going to get us to that, help us with that troubleshooting rather than something's broke. I got to log into a bunch of boxes and do a bunch of diagnostics. No, no, use the script and gather all that information to save yourself a bunch of time. What you're describing, that's just the beginning. You've got agents that are doing that, and then a whole lot more, including testing within a digital twin, digital twin by your definition, a replica of my network that I've stood up in Container Lab and then going from there.