Cisco’s Outshift is pitching the Internet of Cognition as a semantic layer for multi-agent AI systems, meant to let specialized agents share intent, context and reasoning instead of passing data around while a human makes the real decisions.
Vijoy Pandey, senior vice president and general manager of Outshift by Cisco, describes the missing piece as the “connective tissue” that lets separate agents operate as one team. His example is healthcare: one agent handles symptom assessment, another scheduling, another insurance and another pharmacy. They can exchange information today, according to Pandey, but they do not yet coordinate patient care on their own.
Outshift’s proposed stack has two layers. The lower one, called the Internet of Agents, is a connectivity layer for agent discovery, identity and message exchange across domains. The layer above it, the Internet of Cognition, is supposed to give agents shared goals, shared memory and a way to reason through trade-offs together.
What is the Internet of Cognition?
The Internet of Cognition is Outshift’s name for a semantic coordination layer that lets AI agents align on intent, use common context and reason as a group. In plain terms, it is a proposed control plane for agent teamwork, rather than another bigger model.
Pandey argues that the AI industry’s familiar scaling path, training larger models with more data and compute, has produced more capable individual agents. He says the next scaling problem is horizontal: getting agents across different companies, systems and platforms to solve tasks together.
That claim lands in a messy field. Multi-agent systems are already being tested in software engineering, drug discovery and scientific simulation, but results remain uneven. One study cited by Outshift found failure rates ranging from 41% to about 87% across seven open-source multi-agent systems.
Pandey says the weakness is architectural. Connected agents can divide familiar work, he argues, but struggle to hold a common goal and reason toward an answer none of them was trained to produce. He also warns that naive multi-agent setups can do worse than a single agent without a coordination layer.
How Outshift says its agent stack works
Outshift says it has built AGNTCY, an open-source project now under the Linux Foundation, as the connectivity layer for agents to find each other, verify identity and exchange messages using open protocols.
For shared intent, Outshift has created Mycelium, an open-source coordination layer that organizations can clone and test with their own agents. Pandey says internal testing across 14 scenarios found unstructured agent groups reached a decision about one-third of the time, while a protocol requiring agents to declare goals, identify missing information and resolve conflicts raised that figure to 93%. Those are company-reported results, not independent benchmarks.
For shared context, Outshift describes a “cognition fabric,” a policy-governed memory and communications layer meant to stop agents from forgetting useful institutional knowledge after each session. For shared reasoning, the company points to cognitive amplifiers for faster modeling and guardrail technologies for security, cost and compliance constraints.
Outshift also flags the predictable failure modes: agents may delegate tasks unintentionally, accept malicious prompt injections, poison shared memory or gain broader access than their assignment requires.
Its proposed guardrail is Continuous Agent Semantic Authorization, or CASA, an open-source reference implementation. CASA checks what an agent is trying to do and compares each tool request with the original authorized task. In Outshift’s healthcare example, an agent asked to summarize a patient record might try to query an entire database; CASA could deny that request because it no longer fits the task.
Pandey says current controls are usually scoped to a role or session rather than a task, leaving agents with tools they can use too broadly. He also claims that about 90% of the time, an agent has no way to confirm whether it is cleared for the job it received.
For companies testing this idea, Pandey recommends starting with one workflow that crosses three or four teams and currently depends on human approval between handoffs. His suggested metric is whether one agent’s output actually improves another agent’s work, which is a more useful test than counting how many agents were wired together.
More information about Outshift’s proposal is available at Outshift.com. The multi-agent failure-rate study cited by the company is available at arxiv.org.
This story draws on original reporting from MIT Technology Review.
