As the artificial intelligence industry shifts its focus toward multi-agent systems, experts argue that current architectures lack the necessary framework for effective collaboration. While individual AI agents have become increasingly capable, they often operate in silos, unable to coordinate on complex tasks without human intervention. Vijoy Pandey, senior vice president and general manager of Outshift by Cisco, suggests that the industry must transition from vertical scaling—building larger models—to horizontal scaling, which allows agents to function as a unified team.
To bridge this gap, Outshift has introduced the concept of an "Internet of Cognition." This framework relies on a semantic layer that enables agents across different platforms and domains to align goals, share institutional knowledge, and perform collective reasoning. This builds upon a foundational connectivity layer known as the "Internet of Agents," which facilitates agent discovery, identity verification, and secure messaging through standardized protocols. Outshift has released an open-source project called AGNTCY, hosted by the Linux Foundation, to support this interoperability.
The proposed architecture rests on three primary pillars: shared intent via cognition state protocols, shared context through a policy-governed "cognition fabric," and shared reasoning supported by cognitive amplifiers and guardrail technologies. According to internal testing by Outshift, implementing coordination protocols significantly improved decision-making success rates in multi-agent scenarios compared to unstructured groups.
Security remains a critical concern, as multi-agent systems introduce risks such as unintended delegations and memory poisoning. To address this, Outshift developed Continuous Agent Semantic Authorization (CASA), a guardrail technology that validates agent actions against the user's original task in real-time. Pandey advises organizations to begin experimenting with these horizontal architectures by targeting cross-functional workflows that currently rely on human-managed handoffs, allowing businesses to measure how agent insights compound over time.
Source: MIT Technology Review
No comments yet. Be the first to share your thoughts.