Cisco: Network architectures must be rebuilt for agentic AI

Cisco believes network architectures need a fundamental rethink for the artificial intelligence (AI) era. The task is to “modernise everywhere people work,” according to Raymond Janse van Rensburg, the company’s vice-president for networking sales and solutions engineering in Asia-Pacific, Japan and China.

That modernisation must extend beyond traditional offices, as the digital experience is now critical to success in manufacturing (for both people and robots), healthcare, retail, and other sectors.

“We’ve been here before,” Janse van Rensburg observed. Corporate networks evolved with the advent of the internet, changed again when mobility untethered employees from their desks, and shifted once more when cloud computing decentralised traffic flows away from on-premise datacentres.

Now, AI is changing traffic volumes and patterns, so organisations need to rethink their networks, he said. With predictions that each enterprise will deploy thousands of AI agents, it’s important to understand that agentic AI places a higher base load on the network when compared with the bursty traffic associated with AI assistants.

Other pressing issues include the growth of edge computing, the need collaboration between IT and operational technology teams, particularly where robotics are involved, and an overall demand for greater bandwidth.

One of the biggest security risks facing organisations is the emergence of AI-enhanced attacks. When combined with the bigger attack surface that comes with the deployment of AI, the traditional approach to security won’t necessarily work, he warned. Rather, security should be pervasive, and that includes implementing zero-trust security everywhere.

“Identity is the only perimeter,” advised Ray Chopra, senior vice-president and chief security product officer. He argued that identity verification must be woven into every access point with uniform policies, protecting AI agents just as strictly as human users.

Furthermore, organisations must prepare for the post-quantum era by beginning to deploy quantum-resistant cryptography, Janse van Rensburg added.

Compounding these technical challenges is the perennial skills gap. With 64% of organisations predicted to face IT skills shortages by 2026, leaders must identify their requirements early. While external providers can plug the gap, Janse van Rensburg noted the importance of ensuring skills are eventually transferred to internal employees.

All these factors point to the need for a new network architecture purpose-built for the AI era.

Vikas Butaney, senior vice-president and general manager of secure routing and industrial internet of things, said such networks must include provision for agentic operations—allowing the same number of staff to manage increased traffic—as well as built-in security and scalable devices capable of upgrading to 10Gbps and 40Gbps connections.

Cisco’s view is that while cloud-based management is generally suitable, there must be provision for on-premises management for compliance reasons, alongside a hybrid approach encompassing both. This should be achieved with a single set of hardware, one software licence, and one support arrangement.

For example, the Cisco Unified Branch solution provides unified management of Wi-Fi, switches, and secure routers. It offers validated designs for anywhere from 10 to 10,000 branches, with hardware to suit the requirements of each location and a “branch as code” capability for ease of implementation.

While Unified Branch is suitable for organisations with as few as 10 branches – usually implemented via a Cisco partner – Butaney noted it is primarily aimed at enterprises with 100 or more locations.

The system leverages agentic operations through the Cisco AI Assistant to accelerate tasks. This will soon be enhanced by the Cisco AI Canvas (currently in alpha testing), which adds features allowing multiple team members to collaborate on an issue simultaneously.

Regarding security operations, AJ Shipley, Cisco’s vice-president of threat detection and response product management, advised that few organisations can effectively manage their own security information and event management (SIEM). Lacking the personnel to handle the sheer volume of data or the sophistication of modern attackers, organisations should leverage AI to spot patterns and make recommendations.

Chopra pointed out that Cisco has already open-sourced a seven-billion parameter model capable of identifying, for example, when a user action initiates a chain of events leading to a malware download. The company expects to open-source a new, 17-billion parameter model early next year.

AI-based security operations are already delivering results. Shipley noted that a US hospital using Cisco’s technology has reduced threat investigation times from eight hours to just five minutes. According to Chopra, the ability to stop malware quickly has, in healthcare contexts, has saved lives.