Imagine a healthcare system composed of multiple artificial intelligence (AI) agents: one agent managing symptom assessment, a second responsible for scheduling appointments, a third handling insurance claims, and a fourth managing the pharmacy. Each of these agents is an expert in its specific domain. However, each possesses its own distinct knowledge and objectives. Today, while these agents are capable of exchanging data, they are still unable to actually coordinate patient care without a human mediator making decisions to bridge the gaps between them.
"The intelligence is already there. What is missing is the connective tissue that turns four strangers into one team," explains Vijoy Pandey, senior vice president and general manager of Outshift by Cisco. According to an article published in MIT Technology Review Insights, this connective tissue comes from adding a semantic layer—what Outshift calls the "Internet of Cognition"—which enables agents operating across different domains to work together and, critically, "think" together through shared intent, shared context, and shared reasoning.
This semantic layer relies on a connectivity layer beneath it called the "Internet of Agents," which allows autonomous agents to discover one another, prove their identity, and exchange messages across different operational domains. When these two layers are used together, they enable "the next step on the road to distributed artificial superintelligence," says Pandey.
From Vertical Scaling to Horizontal Scalability: The Challenge of Multi-Agent Systems
For years, the artificial intelligence industry has focused on growth. Vertical scaling has led to larger models, trained on larger volumes of data and backed by greater computing power. This process has generated the reasoning and inference capabilities that can serve as the "brain" for AI agents, which are capable of perceiving, reasoning, and acting within digital environments.
While vertical scaling can perpetually produce highly capable individual agents, in order to enable agent-based problem solving across different systems, companies, and platforms, the next axis of expansion must be horizontal, Pandey explains.
Multi-agent systems are currently being explored in fields such as software engineering, drug discovery, and scientific simulations, but their performance so far has been underwhelming. One study found a failure rate of approximately 41% to 87% when evaluating seven open-source multi-agent systems.
"Connected agents handle coordinated action well; they take a task whose shape they have already seen, divide it, and pass it along," explains Pandey. "What they cannot do is hold a goal in common and reason toward a solution for a problem that none of them was specifically trained to solve."
"The gap is architectural, not a prompting problem," Pandey adds. "Without the right coordination layer, naive multi-agent configurations can perform worse than a single agent. The step change will occur when that same team of agents converges on its own to solve a new problem, without the need for a human to stitch the seams between them."
The Connectivity Layer and the Semantic Layer: AGNTCY and the Internet of Cognition
To achieve this goal, Pandey notes that Outshift has built a connectivity layer named AGNTCY, an open-source project now hosted under the auspices of the Linux Foundation. The AGNTCY system allows agents operating across different systems, companies, and platforms to find each other, verify identities, and exchange messages using open, standardized protocols.
As Pandey explains, this allows the "Internet of Cognition" thesis to take a step forward. It creates a semantic layer that enables agents to align their goals (share intent), pool institutional knowledge and cumulative memory (share context), and make collective decisions and trade-offs (share reasoning).
Pandey likens this advancement to the evolution of humanity: "For hundreds of thousands of years, humans became individually smarter, but those gains died with each person who achieved them," he explains. "Around 70,000 years ago, that changed when humans learned to share intent, build cumulative knowledge, and reason collectively. That is the moment scattered individuals became a civilization."
"Agents are now at that exact same threshold. We have built silicon geniuses and granted them agency. What they lack is the layer that allowed humans to become a collective," he says.
The Three Technological Pillars of Agentic Collaboration
Enabling agents to work together collectively rests on three core pillars within the technology stack:
-
Shared intent through cognition state protocols: Cognition state protocols act as the semantic handshake that allows agents to agree on a goal before they act, and subsequently negotiate toward it. Outshift has created an open-source coordination layer called Mycelium, which organizations can clone and deploy with their own agents. In internal testing conducted by the company across 14 different scenarios, unstructured groups of agents reached a decision only about a third of the time (33%). In contrast, a coordination protocol that forces agents to declare a goal, surface missing information, and resolve conflicts before acting raised the success rate to 93%.
-
Shared context through a cognition fabric: A cognition fabric is a shared institutional memory and communication mesh that allows agent insights to compound over time, rather than resetting with each new session. Managed by policies, this context layer solves the issue of "organizational amnesia" by ensuring that the baseline intelligence of the systems only ever increases over time.
-
Shared reasoning through cognitive amplifiers and guardrail technologies: Two types of cognition engines can be utilized together to enable shared reasoning. Cognitive amplifiers accelerate collective thinking and modeling, while guardrail technologies (GATs) establish security, cost, and compliance frameworks. Humans are active contributors to this layer, making judgment calls that the system routes to them (rather than merely reviewing outputs retrospectively after the fact).
Security Risks and Defensive Solutions: Continuous Agent Semantic Authorization
Sharing cognition across multi-agent systems can introduce novel risks, including unintended delegation of authority, malicious prompt injections, memory poisoning, or over-privileged agents with access to permissions and data far exceeding what their tasks actually require. Because of this, environment-specific controls are required to protect against unintended actions or consequences.
"Agents possess human-like attributes but operate at machine speed and scale," explains Pandey. "Everything we have built over the last twenty years—access control, identity, compliance—was built for humans or for machines, not for both together."
The Continuous Agent Semantic Authorization (CASA) protocol—an open-source reference implementation developed by Outshift—is a guardrail technology (GAT) designed to ensure that agent actions remain securely aligned with the user’s original goal through a process of continuous authorization. CASA accomplishes this by reading what the agent is attempting to achieve and then validating each tool request against that specific task.
In a healthcare system scenario, for example, an agent instructed to summarize a patient's medical record might initiate a query across an entire database. This action would prompt CASA to deny the call because the request no longer aligns with the specific task for which it was authorized.
"Today's controls are scoped to a role or a session, rather than the specific task, so an agent granted access to a tool can use it for anything," explains Pandey. "Roughly 90% of the time, an agent has no way to verify whether it is even cleared for the job it was handed."
Practical Recommendations for Enterprises: Starting with Measured Steps
When attempting to scale intelligence horizontally within an enterprise, Pandey advises businesses to begin by experimenting with a single cross-functional workflow that spans three or four teams and currently requires human intervention to authorize the handoffs.
"Establish the process as a small multi-agent system on open, interoperable infrastructure, with a measurable baseline," he advises. "Keep building larger models, layer the horizontal axis on top of them, and shift what you measure. Track where the insight of one agent has improved another agent—that is the signal that the horizontal axis is actually working."
By beginning to experiment now with the intent, context, and reasoning layers, organizations can stay ahead of the curve. "The problems are open, and the infrastructure is still being written as we speak," concludes Pandey. "This is the moment to build it."