A phone won’t join Wi-Fi, and the fix turns out to be a VPN route issue on a firewall. That kind of jump across domains is why network operations feel harder than ever.
In a demo with David Bombal, Cisco’s DJ Sampath showed what Cisco Cloud Control, AI Canvas, and Agentic Ops are meant to do: cut the time between “something’s broken” and “here’s the root cause.” The idea isn’t that AI replaces engineers. It’s that AI keeps context intact while moving across networking, security, observability, and infrastructure.
Once the demo started, that pitch felt much less abstract.
Cisco wants one place to run network ops
Sampath framed the problem well. Networks now span campus, branch, data center, cloud, Wi-Fi, firewalls, identity systems, and AI workloads. Meanwhile, the operator often gets dropped into a pile of dashboards, tickets, and handoffs.
Cisco’s answer is Cloud Control with Agentic Ops. In Cisco’s own launch announcement for its agentic platform, the company describes Cloud Control as a shared workspace for human operators and AI agents. The demo matched that message closely.
The interface looks familiar on purpose. If you’ve used ChatGPT, Gemini, or Claude, you would recognize the layout right away. Cisco wants engineers, operators, and team leads to land in the product and start working without learning a whole new mental model first.
Cloud Control organizes the experience into three groups. At the center are Cisco products such as Meraki, Intersight, Security, and observability tools. To the left are platform services, including inventory and topology. To the right are apps built for jobs that cross product lines.
That structure matters because the data is no longer trapped inside separate tools. Inventory brings routers, switches, and firewalls into one place. Topology adds a live map of how those devices connect. In the demo, Sampath zoomed from a broader view into a London site and then down into the relationships between access points and switches.
The point wasn’t the map alone. It was that Cloud Control brings product context, platform context, and task-focused apps into one control plane.
AI Canvas follows the problem across the network
A plain-language question kicks off the workflow
The live demo started with a problem every network team knows well: “Why can’t G2 get on the network?”
That one sentence kicked off the first agent. It checked the situation and came back with an early read: the Wi-Fi network appeared to be down. Because the issue looked bigger than a quick chat answer, Sampath moved into AI Canvas, Cisco’s shared workspace where people and AI agents investigate problems together.
Cisco has written more about that concept in its Cloud Control and AI Canvas overview. In the demo, Canvas acted like a live troubleshooting board. The original question moved into the workspace, and the assistant suggested the next step instead of waiting for the operator to guess.
That next step was network topology. Cisco added slash commands, much like the ones people know from ChatGPT and Codex, so an engineer can call up a quick network view in the middle of troubleshooting. As soon as the command ran, the topology agent started pulling context.
Sampath made an important point here. The topology wasn’t a static diagram copied from somewhere else. Canvas was talking to MCP servers, reading device relationships, and building a fresh view from live connections.
Topology and specialist agents narrow the fault
The topology view showed the Las Vegas network. Three access points had red indicators, which meant they were offline. Those APs tied back to a remote wireless LAN controller, a Catalyst 9800.
That detail changed the troubleshooting path. Instead of guessing at client settings or blaming Wi-Fi in general, the operator had a tighter fault domain. Sampath sent the context to the assistant on the right side of Canvas, and the troubleshooting agent spun up.
Cisco says that agent uses purpose-built models, including a deep network model and a security-focused model, to reason through the problem. In the demo, the assistant inspected the controller, captured topology evidence, and came back with a new suspicion: the issue might sit in the firewall deployment.
That moment mattered because the workflow crossed domains without losing the thread. One step lived in networking. The next step moved into security. The operator didn’t have to open a new product, copy notes into a ticket, or wait for another team to reproduce the issue from scratch.
The firewall agent then checked the Firepower environment and found the real fault. A site-to-site VPN tunnel was down because OSPF route exchange was missing. In other words, the phone connection problem wasn’t a phone problem at all. It wasn’t even a Wi-Fi problem in the end. It traced back to routing across a VPN path.
In the demo, a client connection issue ended with a root cause of missing OSPF route exchange on a site-to-site VPN.
Sampath said Cisco workflows could then apply the fix with a click. Even if you set the automation aside, the bigger win was speed. A chain of work that could have taken days, several teams, and a stack of tickets took about three prompts.
He also clarified how that happened. Canvas isn’t one giant assistant pretending to know everything. Cisco built a multi-agent system, so the platform can call a topology agent, a Wi-Fi agent, a firewall agent, and other specialists based on the job in front of it.
The Unified Cisco Fabric app links campus and data center
After the troubleshooting demo, Sampath moved to a different problem. More companies now want coding agents and AI workloads to run close to users in offices and campus sites, while still reaching services in the data center. Building that path can take planning, network design, and a fair amount of back-and-forth.
Cisco’s answer is the Unified Cisco Fabric app.
In the demo, Sampath picked a campus called Wayne Tower and a destination called Gotham data center. Then he selected a firewall so the path would stay protected. With a few clicks, the app built the fabric, created the needed VRFs, and mapped the topology from campus to data center with the firewall in the middle.
The Batman-themed names made the demo more fun, but the real message was about time. A job that often turns into a multi-week project was compressed into a short guided flow. That matters for AI projects because the network often becomes the slowest part of getting a new service into real use.
This section also showed why Cisco grouped task-driven apps on the right side of Cloud Control. Some work doesn’t fit neatly inside one product. Fabric setup touches campus networking, data center networking, segmentation, and security. A dedicated app gives teams a cleaner way to complete that work without stitching the process together on their own.
Cisco is treating AI agents as identities that need controls
Once agents start running inside a company, a new question shows up fast: how do you trust them?
Cisco built an agent security app to answer that. The demo brought together zero-trust controls, identity intelligence, and AI guardrails in one view. Instead of forcing customers to line up several products on their own, Cisco presents a single place to see which agents are active, what tools they can call, and which applications they are touching.
That visibility matters because many of these agents rely on non-human identities. If an agent starts using the wrong identity, fails checks, or reaches for resources outside its lane, that can turn into a security problem quickly. The app is built to surface that kind of issue.
Sampath showed a closer look at a Codex agent. The interface exposed the tools it was calling, the subtasks it was running, and the full task life cycle. That gives security and operations teams a plain record of what the agent did, not a black box that only says “completed.”
Policy creation was another key part of the demo. An admin could type a natural-language request and have the system build a policy for a group. For example, a company might roll out Codex to engineering but keep a different setup for non-engineering teams.
Cisco says those mesh policies can be enforced across SD-WAN and at several control points, including secure access, a VPC, a proxy, or a firewall. Behind that one app sit pieces from Secure Access, AI Defense, Secure Client, Splunk, Identity Intelligence, and Duo. The operator sees one policy flow, which is a lot easier to reason about than six product manuals.
Token spend now belongs on the ops dashboard
Security is one side of agent management. Cost is the other.
Sampath didn’t dodge that point when Bombal asked the obvious question: what does all of this cost? Cisco’s answer is another app, this time for agent observability.
Cisco says the app is powered by Splunk and Galileo, the latter a recent Cisco acquisition that the company folded into Splunk quickly. The dashboard shows token consumption, goal progress, and completion data across the agents running inside the business. In plain terms, it acts as a scorecard for your AI workers.
The demo used a vivid example. One autonomous deploy agent had burned through almost 6 billion tokens in the last 24 hours, with a cost near $24,000. That is the kind of number that gets a CIO’s attention in a hurry.
The app didn’t stop at spend charts. Sampath drilled into the agent and showed actions, errors, latency, quality metrics, and performance measurements in one place. That turns token use into an operations problem you can inspect and manage, not just a bill you discover after the damage is done.
Cisco also showed policy controls for token budgets. Just as admins can write access policies in natural language, they can also set spend limits and decide which teams or users get larger budgets. Cisco has talked more broadly about that direction in its post on agentic AI at Cisco Live, and the demo made the idea concrete: agent observability is becoming part of normal IT hygiene.
Open integrations and purpose-built models make the story more believable
A polished demo is easy to dismiss if it only works in a closed lab. Bombal pressed on that point, and Cisco had two useful answers.
The first was ecosystem support. Sampath said Cisco signed up more than 50 partners to join Cloud Control on day one. That matters because few real networks are all-Cisco. In the demo, he showed partner integrations that could pull data into Canvas and make outside systems part of the workflow. ServiceNow tickets and BlueCat IPAM were two examples he named.
The second answer was the model stack behind the product. Cisco isn’t relying on one giant general-purpose model and hoping for the best. Sampath said the platform uses a deep network model trained on networking data, a security model for security work, and a time-series model for data that behaves differently over time. Cisco also uses smaller models for guardrails and calls frontier models such as Codex and Claude when the task fits.
That design also ties into the hallucination question. Bombal raised the concern directly, and Cisco’s response was grounded in data access. Sampath said Cloud Control uses Cisco data fabric, machine data, MCP servers, and APIs to anchor model responses in current operational context. He also said Cisco runs evaluations and ongoing tests to keep the troubleshooting flow on track.
There was one more detail that gave the pitch weight. Sampath said the models draw on more than 40 years of Cisco troubleshooting and debugging data, not only public internet text. He also said more than 50 customers are already using the platform, with more onboarding planned across North America in the coming weeks.
That doesn’t make the hard questions disappear. It does make Cisco’s case stronger. The product isn’t presented as a chatbot that learned networking from random forum posts. It’s presented as an operations system built around live infrastructure data and trained network experience.
Final thoughts
The most convincing part of this demo wasn’t the chat box. It was the way the system carried context across Wi-Fi, topology, controllers, firewalls, routing, policy, and cost without dropping the thread.
If Cisco can keep that quality in real production environments, AI in network operations starts to look less like hype and more like a practical way to cut noise, shorten outages, and help engineers spend more time on decisions that matter.
