When Vijoy Pandey considers the next step for AI agents, his thoughts go back 70,000 years in human history, a critical time in evolution when humans began to really use symbolic language to communicate.

Over hundreds of thousands of years, humans used tools and symbolic information – think primate paintings on the walls in caves – to collect and build knowledge, but were unable to clearly share that knowledge with other humans, according to Pandey, senior vice president of Outshift by Cisco Systems. The knowledge they gathered and the innovation they created were lost to others when they died.

“They did all of this, but it was still a very individualistic intelligence society,” Pandey tells The Next Platform. “There was only power of the individual, there was no power of the collective. What ended up happening about 70,000 years ago, there was the invention of language. Once language happened, instead of being individually intelligent, you had the power of collective intelligence.”

This brought to humans three key capabilities: shared intent and coordination – “instead of one person figuring out what to do, the collective could agree on a common goal and then the collective could coordinate on that common goal,” he says – shared rather than individual knowledge, and collective innovation. Humans could work together to solve any problem.

Fast forward many millennia to today, and Pandey sees a similar situation brewing with agentic AI.

“This is exactly what’s happening with agents,” he says, noting that agents are connecting to each other. “We’re building smarter models, we’re building better agents. We are seeing the first aspects of communication. You are seeing basic collectivity.”

The Need For Collective Intelligence

AI agents can work together, pass messages to each other, and have roles they fulfill, but they need a way to think together “so they operate as a single entity,” Pandey says. “Instead of thinking about AI as a single, individual intelligence, we need to think of it as collective intelligence.”

In a recent blog post and an accompanying 18-page whitepaper, Pandey and others with Outshift – Cisco’s incubator arm focused on such advanced technologies like AI and quantum computing – outlined their vision of what they’re calling the Internet of Cognition, an architecture that addresses a fundamental shift in AI from scaling the intelligence of independent agentic systems to scaling “distributed superintelligence.”

In the introduction to the whitepaper, Scaling Out Superintelligence, Pandey writes that “AI agents, despite their growing capabilities, remain constrained because of semantic isolation. Beyond vertically scaling agents, the next frontier requires new foundational infrastructure for horizontal scaling of intelligence. Such infrastructure would enable shared intent and shared context, accelerate collective innovation, and unlock genuinely emergent capabilities in multi-agent human–AI systems.”

Such systems “can address a far broader class of intelligence problems with greater accuracy and robustness,” he argues, saying that “what we’re doing through this Internet of Cognition is we’re building the infrastructure so multiple agents can come together and think together.”

Building Momentum

The pitch for the Internet of Cognition comes a year after the networking behemoth outlined its vision for the Internet of Agents, essentially creating an open environment where agents – regardless of the vendors that built them – can find each other and work in unison to solve whatever challenges they’re presented with. It also builds on a research paper by Cisco published in in December 2025 that outlined a layered protocol architecture for the Internet of Agents based on a new Layers 8 and 9 for the OSI networking model that is designed for agent collaboration with semantic understanding rather than traditional data delivery between hosts and processes.

The architecture detailed by Outshift builds on what Pandey wrote was “state of the art” for addressing multi-agent systems (MAS) from disparate vendors, organizations, and frameworks that aim to let agents syntactically communicate.

“But as we started deploying MAS in the real world, we realized that the fundamental challenge is not in the syntactic or framework incompatibilities of multi-agent systems, but in the deeper semantic, cognitive, and knowledge layers,” he wrote. “It’s like allowing a Mandarin-speaking Android user to send text messages to a German-speaking iPhone user on iMessages. The frameworks can connect, minimally, but the users cannot semantically collaborate.”

Bring On The Protocols

For this to happen, for this idea of collective agentic intelligence to scale through the Internet of Cognition, the architecture needs to support protocols that address such functions as grounding, discovery, resolution, coordination, and negotiation. Together they “define how agents represent and interpret the meaning of exchanged messages; synthesize and manage collective memory, understanding, and knowledge; minimize coherence problems; resolve between local and collective (global) preferences, values, biases and guardrails,” he wrote.

Pandey pointed to three classes of protocols, including the Latent State Transfer Protocol (LSTPs) for high-fidelity, low-latency communication between AI agent to ensure the “reasoning trajectory” is persevered across endpoints, he wrote. There’s also Compressed State Transfer Protocols (CSTPs) to reduce file sizes before transmission for efficient bandwidth, faster data transfers, and minimal network latency. These are good for low-bandwidth edge or WAN environments.

Also needed are Semantic State Transfer Protocols (SSTPs) to ensure heterogeneous systems can exchange and coordinate data by assuring the semantics – or meaning – of the information is understood.

Beyond protocols, there needs to be a distributed cognition fabric, which Pandey wrote can “provide mechanisms to reconcile state information, but also create emergent context information, and build a consistent worldview with granular enterprise-level policies in place,” and two types of cognition engines. There are cognitive amplifiers to preserve privacy and ensure collective reasoning, along with other help, and guardrail technologies for cost controls, security, and compliance.

Outshift pointed to an example involving its Prometheus networking engineering expert agent and Mythos Corp’s Themis compliance agent agreeing to coordinate – the shared intent – through cognition state protocols, which allow the agents to discovery missing information and resolve conflicting information, and update the cognitive memory fabric.

“Furthermore, these protocols help coordinate and negotiate not just amongst themselves but with the four cognition engines in this deployment,” Pandey wrote. “The cognition fabric provides shared context in the form of working memory, beliefs and design patterns, ontologies like formal network models, and specifics about [one of the] networks, application set, broader infrastructure environment, business goals and service level objectives. The cognition engines help accelerate collective innovation by providing reasoning and formal modeling capabilities. They also provide guardrails in the form of cost, security, and regulation compliance, where innovation can take place rapidly and safely.”

It will add to what Pandey calls the “current state of the art,” pointing to the Internet of Agents that aim for open and interoperable collaboration between agents regardless of vendors, frameworks, and other barriers. He pointed to a number of projects – including the Model Context Protocol (MCP) created by Anthropic for open interaction between AI models and external applications and data sources, Agent2Agent (A2A) for communication between agents, and AGNTCY for discovery, identity, messaging, and observability components – as key initiatives.

He’s expecting the same trend for the Internet of Cognition, writing that “scaling out superintelligence is a call to action. The next era of computation demands a unified commitment to building the infrastructure for distributed superintelligence in an open and interoperable manner.”