I’ve heard from Cisco more than a few times at prior Tech Field Days. As AI continues to penetrate every aspect of IT organizations’ workflows, they offered some interesting insights on the future of monitoring network performance at AI Field Day #8.
When Gladys Isn’t Happy, Watch Out.
When I taught DBAs the advanced Oracle database performance tuning course, I often talked about a mythical user – Gladys – who knew how fast the company’s primary applications should be working better than anyone.
If Gladys wandered by and said “The database is running slow,” you knew you were probably in for a long day and had better pay attention to her complaint … or the next thing you know, your boss would be getting a call from the CIO.
And conversely, if Gladys gave you the thumbs up after a major deployment, you knew you’d done your job.
It appears the folks at Cisco heard of Gladys from their customers – so much so they built a slide around her nebulous complaint.
My point, of course, is it’s relatively simply to diagnose normal, non-AI workloads because network admins, DBAs, and DevOps folks are reasonably skilled at looking at pertinent metrics – I/O wait times, network ping round-trips, application audit logs, average database response time – that quickly point at likely culprit(s).
Nexus: Cisco’s Toolset At the Center of Agentic Operations
AI workloads, however, often defy normal troubleshooting because they span diverse endpoints, including GPUs deployed for intense training, or dedicated compute nodes intended for purely inference operations.
Cisco’s latest strategy for navigating this complex melange is called AgenticOps. At its heart is Cisco AI Canvas, a collection of network-specific models they’ve developed across decades of network monitoring and diagnostic experience, that can deploy an AI agent via MCP to handle AI-centric workload issues.
Cisco demonstrated their AgenticOps strategy concepts through their Nexus Dashboard tools. They showed an AI-assisted diagnosis of networking anomalies related to GPU performance in plain business language, without having to know any of the typical networking CLI commands I’d need to know to drill into what actually was wrong with the corresponding network components and their configurations.
Even better, the Nexus agent framework was able to compare the network configurations against “golden” configurations and recommend potential remediations. As I’m certainly no expert on BGP command line instructions, so I’d probably call in some favors from a network expert colleague before I proceeded to apply the fixes recommended.
I found Nexus offered some intriguing statistics during a demo of it within the Splunk Observability Cloud. Tied to a series of Cisco AI PODs, it’s even possible to view overall “tokenomics” for AI-related operations – for example, how much a single token cost much each POD actively costs in several dimensions, including how many tokens (both input and output) have been expended, cumulative token costs, average cost per request, and any other applicable metrics.
AI Canvas: Human-Agent Collaboration For Issue Resolution
Cisco’s final demonstration showed how their AI Canvas tool made it possible to bring an experienced network engineer into a dialogue with other engineers who were actively attempting to isolate a particular AI workload issue.
The tool was eventually able to help the engineers to draw accurate conclusions about what was actually affecting performance: two GPUs were apparently under-utilized because the network connections to those GPUs’ compute nodes were not working with nominal standards.
Cisco admitted the examples they demonstrated show promise, and they’re planning to expand their Nexus toolsets to their customer base in the future. From my perspective as a definitely network not-know-it-all, I have to admit that even this relatively early stage of their AgenticOps strategy is impressive. At the very least it would keep someone suddenly tasked with figuring out how to solve a complex network issue from acting on minimal intelligence and making an utterly foolish and incorrect decision.
