We recently wrote about the emerging Internet of Agents: A world in which AI agents no longer operate only inside the boundaries of one company or one software platform, but interact with agents run by suppliers, customers, partners and other technology providers.
The basic premise was that this world needs infrastructure. Agents need to identify each other, establish trust, discover capabilities, communicate securely and operate across organizational boundaries. Without that foundation, the agentic enterprise becomes another collection of proprietary silos. But there is a second problem hiding immediately behind the first one: Connecting agents does not mean those agents can actually work together.
Humans know this intuitively. Give 10 people Slack, email and Zoom and you have given them the ability to communicate. You have not created a high-performing team. They may interpret the same goal differently, operate from different information, optimize for competing objectives or simply talk past one another. AI agents have the same problem, potentially at enormous scale.
That is the next frontier being explored by Outshift, Cisco’s emerging-technology incubation group, which has been one of the groups helping build the infrastructure behind the Internet of Agents. Outshift calls this next layer the Internet of Cognition: an architecture intended to move multi-agent systems from exchanging information to forming shared intent, context and coordinated action. And while the term itself is new, the problem behind it is becoming increasingly real.
Much of today’s discussion about multi-agent AI assumes that once specialized agents can talk to each other, collective intelligence will naturally emerge. The early research suggests otherwise.
A 2025 NAACL study created a benchmark specifically to test how LLM agents coordinate. The researchers found that agents performed reasonably well when decisions depended mainly on information in their environment, but struggled when successful coordination required understanding another agent’s beliefs and intentions, with substantial room for improvement in joint planning.
More recent research makes the problem even starker. SILO-BENCH, presented at ACL 2026, tested distributed coordination across multi-agent LLM systems. Researchers identified what they called a “Communication-Reasoning Gap”: agents communicated actively but often failed to turn that interaction into effective collective computation. On the most complex tasks, performance deteriorated dramatically as the number of agents increased.
In other words, more agents exchanging more messages does not necessarily produce more intelligence. That distinction is at the heart of what Outshift is trying to solve. As one Outshift leader put it, “message passing is not collaboration.” Agents may be technically capable of exchanging data while still lacking a shared understanding of what that data means, what they are collectively trying to accomplish or how to resolve competing interpretations.
Consider something as simple as the phrase “Priority 1.” To one organization’s IT agent, that might mean an incident requiring resolution within an hour. To an agent operating in a healthcare environment, it could imply an immediate life-or-death situation. Both agents received the same words. They did not receive the same meaning. As agents increasingly operate across systems, functions and eventually enterprises, those semantic differences become operational problems.
The Internet of Agents work has largely focused on the infrastructure required to make agents interoperable: identity, discovery, messaging, access and observability. Outshift and partners have been developing much of that work through AGNTCY, an open-source effort designed to provide common infrastructure for multi-agent systems. Its goal is to allow agents built on different frameworks and platforms to discover one another, communicate and operate without requiring a custom integration every time two systems meet.
That is a major step. But think of it as the networking layer. Once the agents are connected, the Internet of Cognition asks another set of questions: Do they understand the objective the same way? Do they have enough shared context? Can they recognize that they disagree? Can they negotiate? Can one agent challenge another? Can knowledge discovered by one become usable by the others? And when the system reaches a decision, can a human understand how it got there?
Outshift’s proposed architecture focuses on three areas: protocols that establish shared intent and coordination; a cognition fabric that supports shared, policy-governed context and memory; and cognition engines that can help agents negotiate, coordinate or enforce guardrails.
That may sound futuristic. But Outshift says it is already seeing the coordination failure in its own multi-agent experimentation. Papi Menon, Chief Product Officer and VP of Product Management for Outshift, told us that the team is considerably further into autonomous multi-agent workflows than many enterprises and has begun encountering problems that others have not yet reached. “Agents don’t align and they kind of tend to diverge,” Menon told us.
Outshift has been experimenting with an open-source project called Mycelium, designed to give agents mechanisms for coordination and alignment. According to Menon, in Outshift’s testing, agents without those mechanisms reached alignment in only roughly 36% of cases. With Mycelium introduced to the same agents, he said, alignment rose to approximately 93%. These are Outshift-reported results rather than an independent benchmark, but the magnitude illustrates the problem the team believes enterprises will soon encounter.
“The truth is, when we talk about multi-agent workflows, we are further along on that journey than most people are,” Menon said. Problems such as agents failing to align, coordinate their actions or establish common semantic grounding can sound theoretical, he acknowledged, because many companies are not yet operating autonomous multi-agent systems at that level. That will change quickly if the agent population grows the way vendors expect.
This is not entirely theoretical. Menon described work between Outshift and ServiceNow in which agents operating on separate platforms were able to participate in the same end-to-end enterprise IT workflow. A ServiceNow agent could identify and work a ticket, hand a networking problem to a Cisco agent capable of diagnosing and proposing changes, receive the result back for validation, and then continue the workflow through implementation and documentation. What previously could have required substantial custom integration between the platforms was connected in a matter of days, according to Menon.
The broader market is moving in this direction as well. ServiceNow now supports external AI agents through the open Agent2Agent protocol, allowing agents operating on different platforms to participate in common workflows. That is the Internet of Agents coming to life.
But imagine what happens when the workflow contains not two relatively bounded agents, but 20 autonomous agents representing security, finance, procurement, HR, suppliers and customers, each with different data, goals, permissions and interpretations. At that point, connectivity becomes the easy problem.
There is another reason the cognition layer matters. The future enterprise is unlikely to consist exclusively of agents talking to agents. It will be a hybrid intelligence system of humans and agents reasoning and acting together.
That raises a problem our own research into human-agent collaboration has repeatedly surfaced: humans cannot exercise judgment over systems they cannot understand. Outshift has been thinking about this as well. Its Human-Agent-Experience, or HAX, initiative focuses on how agent reasoning, evidence and actions should be surfaced to people in understandable and consistent ways. The framework emphasizes clarity, user control, recoverability and predictable interaction patterns.
Menon described the challenge this way: even as agents become increasingly autonomous, their activity ultimately has to be “explainable, rationalizable and justifiable to a human.” Human feedback then needs a path back into the multi-agent system. That piece may ultimately be as important as agent-to-agent coordination itself. Collective intelligence is not just about making agents smarter together. It is about creating systems in which intelligence can move among agents and humans without losing context, accountability or judgment along the way.
For CIOs, this suggests that counting how many agents an enterprise has deployed will soon become an increasingly meaningless measure of AI maturity. A company could have 500 highly capable agents and still have very little collective intelligence. The more important questions may be different:
Those questions matter before enterprises reach massive agent scale, not after.
Outshift’s Internet of Cognition is still an emerging architectural thesis, not an established industry standard. Menon himself acknowledged that it sits roughly where the Internet of Agents did a couple of years ago: early enough that many organizations have not experienced the underlying problem and skepticism is understandable.
But that may be precisely why CIOs should pay attention. Technology architectures are easiest to shape before thousands of systems depend on them. The original internet connected computers. The emerging Internet of Agents is beginning to connect autonomous AI.
The next question is whether connecting intelligence can produce something greater than a network full of individually smart machines. As Outshift argued in presenting its work on the Internet of Cognition, this generation may ultimately be less about connecting machines than connecting intelligence. This would provide something much more valuable: agents and humans that can actually think together.