Scaling Intelligence Out: Cisco's Vision for the Internet of Cognition, with Vijoy Pandey

What It Covers

Vijoy Pandey, SVP of OutShift by Cisco, presents the case for scaling AI horizontally rather than vertically — building an "Internet of Cognition" where specialized agents from different organizations discover each other, share context, align on intent, and collaborate autonomously. Cisco's CAPE system (20 agents managing cloud infrastructure) demonstrates this architecture already automating 40% of SRE tasks.

Key Questions Answered

• •Horizontal AI Scaling: The AI industry has focused exclusively on vertical scaling — larger models, more parameters, more compute. A second axis remains largely untapped: scaling intelligence horizontally across networks of specialized agents. Just as language enabled humans to coordinate across distances 70,000 years ago, inter-agent communication protocols could unlock a comparable step-function improvement in collective machine intelligence, potentially compressing AGI timelines more than any single model improvement.
• •CAPE Multi-Agent SRE System: Cisco's internal Community AI Platform Engineer (CAPE) deploys 20 specialized agents managing cloud infrastructure across 100+ tool calls, 5+ user interfaces, and 10+ workflows. Results include a 30% reduction in SRE team workload, full end-to-end automation of 40% of tasks, and response times dropping from hours to near-instantaneous. The system is open-sourced through the Cloud Native Operational Excellence (CANOE) community, with members including Adobe, AWS, and Nike.
• •Four Pillars of Multi-Agent Infrastructure: Before agents can collaborate, four foundational capabilities must exist: discovery (finding agents by capability, not fixed URI), identity and access management, communication (via MCP for agent-to-tool, A2A for agent-to-agent), and observability (monitoring not just infrastructure uptime but agent behavior). The open-source Agntcy project (agntcy.org), part of the Linux Foundation, provides this baseline plumbing for enterprise multi-agent deployments.
• •Task-Based Access Control (TBAC): Traditional role-based access control (RBAC) fails for AI agents because agents shift tasks rapidly and operate non-deterministically. Cisco proposes Tool, Task, and Transaction-Based Access Control (TBAC): agents receive minimum default permissions, privileges elevate ephemerally only for a specific tool call or task, then immediately revert. This zero-trust, least-privilege model is the enterprise requirement for deploying autonomous agents without unacceptable security exposure.
• •Decentralized Agent Directory: Rather than allowing any single platform (e.g., an OpenAI app store) to control agent discovery and identity — and therefore reputation, security, and commerce — Cisco built the Agntcy directory on Distributed Hash Tables (DHTs). No single entity owns the registry. Agents are searchable by capability. Results can return API endpoints or Git code branches. Identity providers support both centralized options (Okta, Duo) and decentralized alternatives, preserving permissionless participation.
• •Two New OSI Layers (L8 and L9): The existing 7-layer OSI network model was designed for deterministic endpoints exchanging data. AI agents are probabilistic endpoints exchanging cognition state. Cisco proposes Layer 8 (syntactic — grammar translation between frameworks like LangGraph and Vertex) and Layer 9 (semantic/cognitive — structured headers that identify communication phase: discovery, negotiation, coordination, or execution). These layers enable enterprises to build deterministic governance on top of inherently probabilistic agent behavior.
• •Cognition Fabric Architecture: Above the protocol layers, Cisco's Internet of Cognition architecture includes a Cognition Fabric enabling many-to-many real-time agent communication with pluggable memory (mem0, BigQuery, or any provider). Stored memory types include ontologies, beliefs, and working memory shared across agent clusters. Cognition Engines sit transparently above this fabric acting as either cognitive accelerators or guardian angels — monitoring for divergence, enforcing compliance, containing blast radius, and preventing the extreme specialization that causes multi-agent systems to silo.