How Cisco and Others Are Preparing f Agent-to-Agent Communication

Imagine a world where your personal AI assistant talks directly to a travel agent AI, which coordinates with a hotel booking AI, which then communicates with a payment agent, all without any human typing a single message. This is not science fiction. It is the near future, and it is called the Internet of Agents.

But there is a problem. Today's AI agents do not truly understand each other. They can exchange messages, but they cannot share meaning. A booking agent might send confirmation details, but another agent receiving that message does not truly grasp the intent behind it. It is like two people speaking different languages with a dictionary but no shared understanding of context.

Cisco wants to fix this. And they are building something they call the Internet of Cognition .

The Problem: Agents Are Talking, But Not Thinking

Right now, AI agents communicate using protocols like MCP (Model Context Protocol) and A2A (Agent-to-Agent). These are useful tools. They let agents discover each other, share messages, and coordinate basic tasks . But according to Vijoy Pandey, who leads Cisco's Outshift incubation team, this is just the beginning.

"These agents are right now just dealing with syntactic communication through MCP, A2A and all of the stuff that AGNTCY provides," Pandey explained to Fierce Network. "But there is no understanding of the semantic connotations behind the payload of that communication" .

In plain language: agents can pass notes, but they do not understand what the notes actually mean. This matters because true collaboration requires shared understanding. If two agents are working together to solve a problem, they need to align on goals, remember what has been tried, and adjust their approach based on what the other has learned. Today's protocols do not support this .

Cisco's Vision: The Internet of Cognition

Cisco is working to build a new protocol architecture called the Internet of Cognition. The goal is to move from syntactic communication (exchanging messages) to semantic communication (sharing meaning) .

Pandey frames it as three big questions his team is trying to answer:

• Can agents share intent, not just instructions?
• Can they share memory and knowledge in context?
• Can they enable collective innovation across agent workflows?

To achieve this, Cisco is building three layers :

Cognition State Protocols

These handle shared intent and coordination. Instead of one agent telling another "run this code," agents would communicate what they are trying to achieve and let the receiving agent figure out how best to help.

Cognition Fabric

This offers the ability to create institution-wide working memories. Imagine a shared space where agents from different departments, sales, support, engineering, can store context and learn from each other's experiences.

Cognition Engines

These may be to speed the pace of innovation or impose guardrails. They are the brains that handle common context and determine what to do.

Why It Matters: The Limits of Current Protocols

To appreciate why where Cisco has stuck its neck out matters, you need to know what is there today. Several protocols have already been implemented :

Model Context Protocol (MCP): A system developed by Anthropic to standardize agent access to external tools and sources of data. It makes it easier for agents to access APIs, databases and file systems.

A2AP, agent to agent protocol (A2A) Google: a developed by allows agents to communicate directly with each other. It takes care of discovery, authorization and task dispatching.

Agent Communication Platform (ACP): Similar to A2A but uses a client-server model with REST APIs and agent registries .

Agent Network Protocol (ANP): Allows agents to communicate directly over the internet using decentralized identifiers and end-to-end encryption .

Agent Payment Protocol (AP2): Also from Google, this lets agents securely handle financial transactions using cryptographically signed mandates .

These protocols solve the "syntactic" layer, they define message formats and transport mechanisms. But none of them solve the deeper problem of shared meaning. As Roy Chua, founder of AvidThink, put it: "The challenge of ensuring agents have genuinely aligned semantic understanding, not just compatible message formats, requires solving deep problems in representation learning and ontology alignment" .

The Industry Is Paying Attention

Cisco is not alone in recognizing this need. In the industry, standards bodies like NCAP are preparing foundations for agent-to-agent communication.

ETSI, the European Telecommunications Standards Institute has active work on specifications for AI agent interfaces and in next generation mobile networks. Their work item, led by Deutsche Telekom, Huawei, China Telecom, and others, analyzes the gaps between existing protocols like A2A and MCP and the requirements for agent communication in telecom environments . This is serious standardization work, not just research.

IETF, the Internet Engineering Task Force, is also exploring these questions. A draft document from China Unicom discusses the impact of AI agents on network infrastructure, examining scenarios for agent communication across local networks, wide area networks, and wireless networks . The document identifies gaps in agent identity management, mobility support, and secure communication channels.

Another IETF draft from Huawei and CAICT proposes the concept of an Agent Gateway, a new kind of network entity that would handle agent discovery, access control, and information distribution . This gateway could be a trusted intermediary that oversees how agents connect with one another, and enables secure communications.

TM Forum, the global industry association for telecom operators, has also been discussing autonomous network standardization. Recent meetings covered work in IETF, ITU-T, and ETSI on AI agents and network management . The conversation is happening across all major standards bodies.

What This Means for Networks

If agents are going to communicate at scale, the underlying network infrastructure needs to change. Cisco is preparing for this with new hardware designed specifically for the "agentic era."

At Cisco Live EMEA in February 2026, the company announced the Silicon One G300, a new Ethernet switch chip delivering 102.4 terabits per second of bandwidth . This chip is designed to handle the unpredictable traffic patterns of agentic inference, where autonomous agents continuously interact across distributed environments.

Brendan Burke, Research Director at The Futurum Group, explained: "As AI workloads shift toward agentic inference, where autonomous agents continuously interact across distributed environments, the network must handle unpredictable traffic patterns, unlike the structured flows of traditional training" .

G300 has fully shared packet buffer architecture which absorbs the congestion from any ports that can provide for up to 33% increase in network bandwidth without adding more physical capacity. It also has passive packet processing capability that enables one switch to process diverse type of traffic for both training and inference concurrently.

Beyond hardware, Cisco is building software for the agentic era. The company recently announced AgenticOps, an agent-first IT operating model for autonomous action with built-in oversight . New capabilities include autonomous troubleshooting, continuous optimization, and trusted validation, all designed to let AI agents manage networks while keeping humans in control.

At the application layer, Cisco introduced the AI Canvas for data center networking, which lets operators interact with the network using natural language and visualize complex multi-tiered flows. This is the interface between human intent and agent execution.

The Challenges Ahead

Building an Internet of Cognition is not easy. Roy Chua points out several hurdles :

• Model differences: Different AI models have fundamentally different architectures. Making them share meaning is technically challenging.
• A question of security: Agents with sensitive data and credentials must have strong protectiveness measures.
• Intellectual property: Companies may not want their representatives disclosing too much of how they operate.
• Coordination problems: Even with shared context, multi-agent coordination remains challenging.

Chua advises telcos and enterprises to watch these developments but avoid committing significant capital until specifications mature. "We'll likely know in 1-2 years whether this becomes a foundational standard or an effort like intent-based networking, which until recently, couldn't meet its own stated goals," he said .

The Road Ahead

Cisco predicts that AI will add two new layers to the traditional seven-layer communications stack :

• Layer 8: Syntactic protocols for agent communication (what exists today)
• Layer 9: Semantic cognition protocols for shared meaning (what Cisco is building)

"That implies that every networking stack moving forward will have to deal with not just Layer 4, not just Layer 7, but also the probabilistic communication at Layer 9," Pandey said .

For enterprises building with AI agents, the message is clear: prepare for a world where agents don't just talk, they think together. The standards are being written now. The hardware is being built now. And the organizations that pay attention to these developments will be the ones best positioned to benefit when the Internet of Cognition arrives.