Cisco touts ‘Internet of Agents’ for secure AI agent collaboration

When AI agents begin to proliferate, a new, open structure will be needed so they can securely communicate and collaborate together to solve complex problems, suggests Cisco.

To head off potential problems such as agent sprawl and fragmented infrastructure, Cisco’s advanced research outfit Outshift is proposing the “Internet of Agents,” an open-sourced, three-layer architecture that would enable quantum-safe, agent-to-agent communication to allow AI agents to collaborate autonomously and share complex reasoning, according to Vijoy Pandey, senior vice president of Outshift by Cisco.

“We’re heading into a world of ‘agent heterogeneity’– different vendors, different capabilities, minimal standardization – which will create a growing challenge: agent sprawl,” Pandey said.

Today’s infrastructure was built for humans sharing content like websites, emails, and videos. But AI agents are fundamentally different, Pandey said. “They don’t just share information; they collaborate, reason, and take autonomous actions in real-time. More importantly, they introduce probabilistic computing into our technological foundation,” he said. 

“As AI gets built into every application and service, organizations will find themselves managing hundreds or thousands of discrete agents. Without open standards and frameworks, this diversity creates chaos,” Pandey said. “It’s like the early days of networking – we need common protocols and standards so these agents can discover, communicate, and collaborate with each other effectively. This standardization and interoperability will be essential for enterprises to effectively manage and scale their AI initiatives.”

In a blog about the need for an Internet of Agents, Panday cited a real-world enterprise IT example:

“In enterprise IT, deploying a sales forecasting SaaS platform requires collaboration across multiple AI agents. A planner agent decomposes business requirements into specific build tasks, while ServiceNow agents coordinate with Salesforce and Hubspot systems for deployment. Infrastructure agents from Cisco and Microsoft simultaneously validate security, identity and access, costs, and SLO compliance. This all requires sophisticated agent-to-agent reasoning and without open standards, each of these interactions becomes a custom integration project.”

“This is a pretty straightforward example, but that’s what we’re getting into. And nobody is looking at that agentic communication problem,” Pandey said.

IBM and Microsoft looking at agent overload

Cisco isn’t alone in exploring the impact of the proliferation of agents. For example, Microsoft has talked about how AI agents will impact application development.

IBM, too, in a recent report wrote: “Agentic AI is quickly transforming the role of individual contributors. As simple AI assistants are supplemented by AI agents with more advanced capabilities, employees will need to manage entire teams of agents that are completing tasks autonomously—and learn how to work with chat-based supervisory AI agents that can help streamline this process.

“This ‘third wave’ of AI promises to transform workflows wholesale.7 In fact, nine in 10 executives now say their organization’s workflows will be digitized with intelligent automation and AI assistants by 2026—and 77% of executives believe gen AI will enable connected assets to make autonomous decisions by 2026. Executives also report that the volume of decision-making by digital assistants will increase by 21% in the next two years due to generative AI. This will have huge implications for operating models, as organizations must create new structures that give employees oversight over autonomous decision-making—and manage the new risks it creates,” IBM stated.

Outshift’s vision for Internet of Agents

The Internet of Agents will address these challenges, Pandey said. Like the internet and cloud, the Internet of Agents that will be built on three interconnected layers that Outshift defines as:

AI-native agentic applications: This layer encompasses the full spectrum of agentic applications—from business workflow automation to scientific discovery to social interaction. Think of it like a movie production, where specialized teams (writers, actors, cinematographers, editors) collaborate to create something greater than any individual could achieve. Similarly, AI agents will specialize and collaborate across domains, from software development to drug discovery to embodied agentic robotic workflows, etc.

Agent communication platform: This layer provides the fundamental protocols and standards for how AI agents discover, authenticate, and interact with each other. Like TCP/IP for the original internet, these open standards must enable any agent to seamlessly participate in the network, regardless of its creator or purpose for both hardware and SaaS products.

AI and quantum-safe infrastructure: This foundational layer delivers the secure, scalable infrastructure that enables all AI agent interactions. It combines high-performance computing and networking with quantum-resistant security built in from the ground up. Through quantum networking capabilities and advanced security protocols, this layer ensures that agent communications remain protected against both current and future quantum threats. This proactive approach to quantum safety creates a trusted foundation for the entire agent ecosystem to build upon.

A major challenge with enabling open communications is that agents don’t talk via straightforward API, Pandey said. 

“Yes, they might use APIs in the back end, but the payload that is going between agents is highly probabilistic in nature. You’ll say things like, ‘find me the best x.’ What is best? It’s not a very well-defined thought. So, agent X might think best differently from agent Y, so probabilistic input and probabilistic output. So the agent might come back and say, ‘yes, this is the best x with 90% confidence.’ So there’s probabilistic input, probabilistic output,” Pandey said.

Multi-modal information is being exchanged, Pandey explained. “You’re exchanging not just APIs across the internet. You’re exchanging audio, video, texts, like science data, like protein folding data. You’re exchanging really massive sets of data between agents,” Pandey said.

The Internet of agents layer can assure that they’re actually doing what they’re supposed to do, he said.

Pandey said Outshift is working with some partners for the Internet of Agents already, but declined to name them. But like Linux, Kubernetes, SONiC and other open-source technologies, the Internet of Agents will feature open-source community development, Pandey said. 

“Going forward, when people develop agents, we want them to speak Internet of Agents right away,” Pandey said. “It’s like saying, if you have the BSD TCP/IP stack on your host machine, the machine is already talking IP. So when you come out with an IP router, voila, it just works. So that’s the notion that we want to at least work with.”