Cisco Outshift proposes semantic layer for multi-agent AI coordination

Vijoy Pandey, senior vice president and general manager of Outshift by Cisco, has outlined a framework for distributed artificial superintelligence that relies on a semantic layer termed the Internet of Cognition. The proposal aims to enable autonomous AI agents across different domains to coordinate, share intent, and reason collectively, moving beyond the industry’s long-standing focus on vertical scaling of model size to horizontal scaling of agent networks.

The framework addresses the limitations of current multi-agent systems, which Pandey describes as lacking the connective tissue required to function as a unified team. While individual agents can exchange data, they often struggle to coordinate complex tasks without human intervention. Outshift argues that the gap is architectural rather than a matter of prompting, noting that without a proper coordination layer, naive multi-agent setups can perform worse than a single agent.

To facilitate this coordination, Outshift has introduced AGNTCY, an open-source connectivity layer hosted by the Linux Foundation. AGNTCY allows agents across different systems, companies, and platforms to discover one another, prove identity, and exchange messages through open, standardized protocols. This connectivity layer underpins the Internet of Cognition, enabling agents to share intent, context, and reasoning.

Internal testing cited by Pandey suggests that using the Mycelium coordination protocol, which is also open-source, raised multi-agent decision-making success rates from approximately 33% to 93%. Mycelium functions as a cognition state protocol that requires agents to declare goals, surface missing information, and resolve conflicts before acting, thereby creating a shared intent across the network.

Outshift also introduced Continuous Agent Semantic Authorization (CASA) to manage security and compliance risks associated with agent interactions. CASA is an open-source reference implementation that ensures agent actions remain aligned with the user’s original goal by checking each tool request against the authorised task. This approach addresses the risk of over-privileged agents by denying calls that no longer match the specific task intent.

Pandey likens this progression to the development of human civilisation, suggesting that around 70,000 years ago humans learned to share intent and reason collectively, transforming scattered individuals into a society. He advises businesses to begin experimenting with one cross-functional workflow that spans three or four teams and currently requires human authorisation for handoffs, using open, interoperable infrastructure to measure the benefits of horizontal scaling.

The framework rests on three pillars: shared intent through cognition state protocols, shared context through a cognition fabric that prevents organizational amnesia, and shared reasoning through cognitive amplifiers and guardrail technologies. Outshift contends that this architecture is essential for enabling agentic problem solving across different systems and platforms, marking the next step on the road to distributed artificial superintelligence.