Cisco Says AI Agents Broke the Enterprise Network in February

AI agents consumed more tokens than humans for the first time this February, Cisco's president for Asia Pacific, Japan, and Greater China, Ben Dawson, told the opening keynote at Cisco Live APJC in Melbourne this week. Agents now consume tokens at roughly seven times the rate of humans. "Each agent is generating vastly more traffic and data than a human performing the same task," Dawson said.

The number anchors Cisco's central claim at the event: networks built for people will not hold up when thousands of agents work around the clock. After years of AI conversation centered on GPUs, AI factories, and data centers, Cisco shifted the focus to the campus, the branch, and the connections between them.

Why does agent traffic break existing networks?

The chatbot era produced bursty, mostly north-south traffic. A prompt went to the cloud; an answer came back. Jeff Schultz, Cisco's SVP of portfolio strategy and marketing, explained that agents change the pattern. They still push heavy north-south traffic to reach models and context, but they also generate large volumes of east-west traffic as agents talk to other agents across desktops, campuses, branches, and data centers.

"All of this again comes back to the idea of needing to fundamentally reimagine the architecture for networking in the campus and branch," Schultz said.

Dawson also flagged a wired-network renaissance. "AI is going to put new demands on wired infrastructure, making it even more critical," he said. Many organizations have treated the access layer as a commodity for a decade. Agents will test that decision.

How big is the readiness gap?

Oliver Tuszik, Cisco's chief sales officer, cited IDC data showing that only 9% of organizations in the region believe their networks are fully ready for new AI and digital initiatives. He also cited research indicating that 85% of organizations globally are experimenting with agents, but only 5% have deployed them to production. In conversations with CIOs, the network is often what keeps them stuck in that gap.

What did Cisco announce?

Cisco's answers span its entire stack:

• Data center, scale-out: 102.4 Tbps systems based on Silicon One G300, with air- and liquid-cooled options and a choice of NX-OS or SONiC.
• Data center, scale-across: the deep-buffered Silicon One P200, connecting data centers across metro, long-haul, and transcontinental distances.
• Campus telemetry: ThousandEyes integrated into Wi-Fi 6E and Wi-Fi 7 access points at no extra cost for customers with eligible Networking Essentials and Advantage licenses. Schultz conceded that ThousandEyes has been "a little bit hard and difficult to license and deploy" and is now making it a default.
• Operations: Cisco Cloud Control, the unified operations layer Schultz called the "single front door" to the portfolio, combining cross-domain telemetry, purpose-built models, and Cisco and third-party agents.

Two operational details matter to network engineers. First, Schultz said no new Cisco product will ship unless it is integrated with Cloud Control. Cisco historically built products in silos and left integration to customers; under Chief Product Officer Jeetu Patel, the company has committed to absorbing that complexity itself.

Second, availability is limited. Cloud Control currently runs only on US-based infrastructure. Australia, the EU, and the UK are scheduled for February 2027, with Japan, India, and Canada following in May.

Putting assurance data on every new access point also gives Cisco's AgenticOps agents something real to work with. Agentic operations depend on the telemetry beneath them.

How does Cisco propose to secure agents?

Schultz told the audience that "the line between observing an agent and securing an agent is beginning to blur." He cited an agent that suddenly deleted 10,000 files. Was it compromised, or did it decide on its own to free up disk space? Neither a security tool nor a monitoring tool can answer that alone.

Patel has described the framework as a "continuous trust loop" and has argued that AI cannot scale without security, governance, and observability. It has three parts.

Testing. Cisco AI Defense is adding Anthropic's Claude Mythos to its adaptive red-teaming engine. Cisco set firm limits: Mythos will test only targets the customer owns and has registered; each engagement is cryptographically tied to a named operator; sessions are time-bound and can be stopped; and customers cannot prompt the model directly. It arrives soon for select enterprises, with broader availability for qualified customers later.

Cost and quality. Splunk Agent Observability's Tokenomics feature pairs each agent's spending with evaluations of its work, using Amazon Bedrock and CloudWatch data. It enters private preview in December, with general availability on AWS Marketplace in early 2027. Splunk's new Eval-G1 small language models run evaluations and guardrails at up to 98% lower cost than a commercial LLM, making it practical to check all agent traffic rather than a sample. Schultz argued tokenomics is a security issue too: in agent-driven attacks, the side that spends tokens more efficiently wins.

Network-security integration. Cisco IQ vulnerability, lifecycle, and configuration data now flow into the Splunk Platform via Cisco Data Fabric, letting a SOC analyst or an AI agent match a live detection in Splunk Enterprise Security to the exposed switch or router. New MCP and ServiceNow connectors convert findings into asset-linked problem records. This closes a gap between network state and security workflows that has slowed incident response for years.

Speed drives the design. Jacqui Guichelaar, who leads Cisco's customer experience organization in APJC, said the share of exploits occurring on or before disclosure day has risen from 22% in 2019 to 78% this year. Cisco also expanded Live Protect, which pushes "shields" that block exploitation of critical vulnerabilities until a patch is applied.

What did customers report?

John Kelvie, chief digital and information officer at Asahi Beverages, described a greenfield Melbourne head office for roughly 1,000 employees, built on Cisco's SD-Access architecture so the network could be segmented and secured from the start. "We couldn't build tomorrow's AI organization and enterprise on yesterday's technology," he said, calling the network "the nervous system" of the company. Asahi has since brought OT on the production floor onto a single IP network, aiming for a consistent experience whether an employee sits at the head office, a brewery, a distribution center, or in Malaysia or New Zealand. The technology, Kelvie said, "just had to work."

Asahi already uses AI for demand sensing, forecast optimization, trade promotion, and back-office automation. Kelvie said the next phase shifts from team-level experiments to "fewer, bigger initiatives" across the enterprise. His biggest obstacle is data, not people or technology. "AI is going to amplify your strengths, but it's also going to amplify your inconsistencies," he said. Asahi still runs separate processes in parts of its alcohol and non-alcohol businesses, creating different versions of the truth. Harmonizing them is a prerequisite for the fully automated distribution centers the company is building, which promise "one customer, one order, and one delivery."

Yessie Yosetya, CIO of Indonesian operator XLSmart, inherited two networks after the merger that created the company: one scoring 3.2 on a NIST-based maturity scale, the other 1.6. The easy option was comparing the two legacy stacks tool by tool, 54 times over. Instead she redesigned around five areas: monitoring and detection, identity, network, applications, and an agentic SOC. "We are not choosing two yesterdays," she said. The result cut the security tool count from 54 to 14.

Her agent experience is a warning for teams building on their own. XLSmart put eight agents into production for red, blue, and purple teaming, then found it could not scale them without standardization and guidance. "We cannot beat machine speed with human speed," she said. Her advice on partners: watch how a vendor behaves when a problem arises. The one that focuses on finding a fix is the one that can execute.

Does one vendor hold the whole stack?

Cisco's argument is that silicon, systems, optics, operations, security, and observability perform better when they come from a single vendor. Customers concerned about lock-in will be skeptical, and the open-ecosystem story still has to prove itself. The thesis still holds: agents will change traffic patterns, multiply non-human identities, and shrink response windows faster than most network teams expect.

For IT leaders, the practical steps are concrete. Revisit campus and branch capacity plans with east-west agent traffic in mind. Treat telemetry as a design requirement, not an add-on. Bring network, security, and observability teams together now, because no single team can decide alone whether an agent's behavior is a bug or a breach. With only 9% of organizations believing their networks are ready for AI, getting the network ready may matter more than which model an enterprise picks.