Large software companies are releasing more and more AI agents of their own, and IT departments have to somehow manage them all. Cisco therefore wants to secure AI agents from the network, which, according to Frank De Reymaeker, sees how agents talk to each other.
“It is often described as the Wild West,” says Frank De Reymaeker, Director Solutions Engineering at Cisco Benelux, about AI agents in companies. “Mainly because it is not always clear to IT and business leaders how to get started with AI agents and how to tackle the various challenges around infrastructure, cybersecurity, costs, governance… you name it.”
According to De Reymaeker, pretty much everyone already uses Copilot in Outlook and Word. Most companies are still at the start with AI agents. He expects their number to grow quickly. “Tomorrow, every company will have ten, fifty, a hundred agents. And every employee will probably have agents of their own too.”
Where the challenge lies
Do AI agents really differ that much from familiar AI chatbots such as ChatGPT and Gemini? “Absolutely,” says De Reymaeker. “AI agents build on AI models, but they can carry out tasks autonomously or semi-autonomously. Think of replying to your emails, starting tasks and processes, or solving complex problems together with other AI agents. You can therefore see them more as full-fledged digital colleagues who work alongside human employees and carry out tasks on their behalf, with access to many different systems.”

That calls for a new approach to network management and to security. Scalable capacity is needed for the needs of employees, devices and now also a multitude of AI agents. For IT teams, the operational management of people, things and agents is no easy task either. They also want to guarantee reliable use of AI with a tight approach to cybersecurity . In other words, no rogue agents.
Securing AI agents from intake
AI can indeed only be truly used if you can use it safely. Do you currently have visibility into which AI tools are used within your company? And if you also use AI agents, are you sure they run on a reliable model?
To that end, Cisco continuously scans the available models and gives them a safety or reliability score. “That is step 1, it is really like the intake interview.” It has to show whether the model is reliable, “or actually a repackaged attack model, which we often see.”
Then come the guardrails around the model and the agent. Cisco determines which identity and which rights an agent receives, and monitors whether its behaviour stays within the guidelines the company has set. In this way, Cisco wants to protect agents from people who want to make them do something wrong, and protect companies from agents that run amok. According to De Reymaeker, there have been plenty of examples of the latter in the news in recent months.
Zero trust for agents
De Reymaeker starts from a lesson learned over twenty years of security. “If you only place security in certain spots, there are always ways around it. If you build security in everywhere, it becomes hard for hackers to get in.” With AI agents, the network gets a role of its own in this. “Agents talk to each other. The network sees all of that.”
He says some companies view AI security too much as a separate domain. Cisco links AI security to network security, and that starts with zero trust. That principle first applied to people and later to people and things. “Today, we very clearly look at zero trust for people, things and agents.”
Agents talk to each other. The network sees all of that.
Frank De Reymaeker, Director Solutions Engineering at Cisco Benelux
According to De Reymaeker, with the addition of things, the number of points IT has to manage already rose from one to three to five. Agents could multiply that by another five. According to De Reymaeker, the network constantly monitors what an agent does with its rights. “Are you talking at certain moments to things you are not supposed to talk to? Are you trying to break out of the environment we have provided for you?”
With the acquisition of Splunk in 2024, Cisco brought in knowledge of SOCs and observability. Cisco is now turning that into an Agentic SOC in which agents work alongside security teams, and into Splunk Agent Observability, which evaluates agents and models and shows what agents do in an environment, “both within their model, within their thinking, but also within their communication.”
Cloud Control as moderator
De Reymaeker sees a third wave here, after software defined and after automation, in which companies manage their environment as code. Cisco calls this wave AgenticOps. In it, the administrator gets help from agents that analyse all traffic, all patterns and all incoming data, give advice and propose themselves how they would tackle a problem. That should allow the environment to be adjusted faster, including for security and defence against cyberattacks.
The platform for this is Cisco Cloud Control, which Cisco introduced in June at Cisco Live in Las Vegas. De Reymaeker calls it “not really a product”. Cloud Control connects the domains of Cisco and of other vendors, splits a problem into sub-problems and looks in each domain for the information that is needed.
He uses the image of a table. IT departments already tried to get the teams for servers, security, applications and network around the same table. Now the SOC with its agents and Cloud Control become the moderator of that conversation. “Splunk, give me the application information. Server, give me the server information. Network, do you see anything happening? And ThousandEyes, do you see anything happening in the cloud that might be affecting it?”
Cloud Control also works with platforms from other vendors, including ServiceNow. If someone creates a ticket in ServiceNow, that ticket can ask Cloud Control to investigate the problem. “Bring everyone around the table,” says De Reymaeker.
What an agent costs
As the use of AI, and especially the use of smart AI agents, increases, there is also more and more discussion about the associated costs. According to De Reymaeker, that discussion is “still firmly in the storming phase. Everyone is searching.” Organisations want a grip on the cost and the volume of tokens consumed. For De Reymaeker, that tokenomics discussion is already in a second phase. First it was about the cost of intelligence, the price per token. Now it is about the cost per delivered value. “The right model for the right problem at the right cost.”
Each of the 90,000 employees at Cisco has an agent of their own.
Frank De Reymaeker, Director Solutions Engineering at Cisco Benelux
At Cisco itself, according to De Reymaeker, each of the 90,000 employees has an agent of their own, and everyone sees their own token consumption. The solutions Cisco offers customers give insight into token consumption and token costs. They also give customers the option to run the right model in the right place, so they keep this cost under control.
A visible acceleration
The market is changing fast, says De Reymaeker. In less than a year, the market evolved from the first use of AI tools, to defining an AI strategy, to exploring the use of AI agents and in-house SLM models. The more internal data, intellectual property or sensitive data there is, the more organisations look at on-prem solutions.
According to De Reymaeker, an evolutionary approach is the best. “The smartest innovators simply started building, but with a number of fundamental principles for scalability, resilience and security. The applications then grew gradually.”
At Cisco, employees get all the tools to experiment with AI, in an environment that is securely shielded. De Reymaeker sees that same role for the IT department at customers. “Make it possible, enable it, but enable it in a secure way.”
What role integrators and MSPs play in this, and how they charge for it, is still open. Cisco closes the keynote of Cisco Connect Benelux with a conversation with some of its platinum partners about how they see their role evolving. It is the first Cisco Connect that Cisco organises for customers and partners as a Benelux organisation.
Cisco Connect Benelux takes place on 26 November at the Klokgebouw in Eindhoven, with an IT Leadership Track, technical sessions and a World of Solutions with eighteen Cisco demo stands. You can register via the Cisco Connect Benelux registration page.
