IT leaders test the waters of autonomous network operations

NetOps agents fix overlooked QoS issues

So far, Danfoss has put NetOps agents into production at 20 or 30 of its hundreds of data center sites, with plans to scale up, though the network operations team has been absorbed in other projects related to M&A by the company recently, Baastrup said.

In addition to using agents for repeated, toilsome network operations tasks, the company's NetOps team has reaped unexpected benefits from the Dynamic Capacity Optimization agent introduced by HPE in May.

We have realized all the things that we don't get to.

That agent works at a more subtle level than outage troubleshooting. Instead, it autonomously identifies capacity bottlenecks and tunes networks to improve end-user capacity, coverage, and roaming experiences for wireless users, even when services are still operating.

Using it has been an eye-opening experience for Danfoss network operators, according to Baastrup.

"We have realized all the things that we don't get to," he said. "Where an access point is overloaded, or it should have been restarted, or there are two access points that have interference because somebody put a metal wall somewhere in the warehouse, and nobody actually told the networking department."

As to why end users hadn't reported slow or unstable network performance, Baastrup speculated that they might have grown accustomed to degraded service,  particularly after the shift to remote work during the COVID-19 pandemic.

"People will call us when it doesn't work," he said. "But when it just slows down or gets a little unstable, maybe people don't call us."While the agents have revealed unexpected benefits, Danfoss is moving cautiously in giving them more autonomy. Limiting the number of sites where the agent is rolled out and co-piloting with a human operator at first are two ways the company has mitigated the risks of rogue or misaligned agents on its networks. Otherwise, the change management and release process has been the same as for other software rollouts – and employee onboarding -- at Danfoss.

"We actually treat agents like humans to a certain extent," Baastrup said. "We give them dedicated permissions and rights, just like we would do for a human being. And you would never give a student your access rights when they start working, or say, 'Oh, you can just use my login for all of this.'"

Troubleshooting agent eases alert fatigue

Cisco customer Room & Board, a furniture retailer in Golden Valley, Minn., has begun using the Cisco AI Assistant agentic tool with its Meraki wireless networking dashboard to collect, correlate and report regularly on system alerts through a Teams channel for WebEx systems managers.

"Every couple of hours, we get a summary of what's been happening, if there's anything that the AI was able to discover for potential troubleshooting or if whatever issue was seen was resolved, basically doing what I would call first-level triage of events," said Mark Rodrigue, senior network engineer at Room & Board. "Is this something that's happened before? Does it keep happening? And, giving us the next steps: is this something we need to pivot to address further, or can we dismiss it as normal?"

Next, the Room & Board team will start allowing the troubleshooting agent to collect data on incidents from more systems in the environment, and expand its agent management using the new Cisco Cloud Control interface.

"We are not at the point yet where we're doing agent-based remediation," Rodrigue said. "Ideally, we want to get there, but the initial triage and data collection is super valuable and time-saving."

Agentic incident triage is much faster and involves much less manual effort than parsing long files using the company's Splunk log analytics system, as was the traditional method, he said.

Another Splunk user, speaking during a breakout session at the Cisco subsidiary's .conf event in September, also found an AI SRE agent an improvement over previous generations of machine-learning-based AIOps tools, even though it doesn't yet take autonomous actions on infrastructure.

"You're asking a question, and it's pulling all these datasets together on the fly -- from my perspective, I would love for it to be able to take action because that saves us time," said Van Wolfe, senior vice president of platform engineering at Repay, an electronic payments provider in Denver, during Q&A at the end of the .conf session. "Most of the experience that I've had with machine learning, you've got to do a lot of customization, so I see [agents] as a more intelligent way to get to that same kind of outcome."

Cisco's September AgenticOps report cited testing showing that AI agents can generate up to 450% more network traffic than conventional applications, adding complexity that Cisco predicts will further drive up autonomous network adoption. But for now, Room & Board has enough complexity in managing distributed hybrid cloud resources to warrant agentic troubleshooting help, Rodrigue said.

"It's not only getting complex in that the amount of services and providers that are involved, but the speed at which they're involved," he said. "Even just looking at our [existing] application stack as it continues to evolve and spread – across hyperscalers, across instances within hyperscalers, across separate providers within different hyperscalers – there's just so much variability in all of it, and none of it you have direct visibility into."

Reaching autonomy will take much more time for Rodrigue's team, which is still getting used to interacting with agents, he said.

"We see agentic AI as a force multiplier; eventually, it will become closer to a coworker, hopefully, effectively doubling the size of our team someday," he said. "But there's a lot that comes with that. There's got to be guardrails. There's got to be documentation. There's got to be audit trails. There's got to be trust that's continuously earned."

Beth Pariseau, senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism. Have a tip? Email her or connect on LinkedIn.