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

Episode Summary

Executive Summary: Vijay Pandey argues that frontier AI needs an “internet of cognition”: a distributed, permissionless layer for agents to discover each other, establish identity and trust, share context, and coordinate safely. Cisco’s Outshift is building this via Agency and internal systems like CAPE, showing how multi-agent systems can automate enterprise workflows while preserving least-privilege access, observability, and governance.

Main Topics: Cisco as a distributed systems company (Priority: 5/5): Pandey frames Cisco’s core identity as moving information securely and reliably, emphasizing networking, security, observability, and collaboration as the foundation for scale-out computing and AI infrastructure. From the OSI stack to new AI layers (Priority: 5/5): The conversation revisits the classic networking stack and argues that AI requires two additional layers: syntactic communication (layer 8) and semantic/cognitive communication (layer 9) to support intelligence-to-intelligence interaction. The Internet of Cognition vision (Priority: 5/5): Pandey proposes a horizontal scaling paradigm for AI where specialized agents collaborate across vendors and domains, sharing intent, context, and memory to solve novel problems without human intervention. Agency project: plumbing for multi-agent systems (Priority: 5/5): Cisco’s open-source Agency project provides discovery, identity/access, communication, and observability for agents, enabling them to connect and operate in enterprise environments. Identity, trust, and least privilege for agents (Priority: 5/5): A major focus is replacing human-centric RBAC with tool/task/transaction-based access control (TBAC), so agents get only ephemeral, minimum necessary permissions tied to specific actions. CAPE/Jarvis and enterprise automation (Priority: 4/5): Cisco’s internal multi-agent system for SRE work automates repetitive operational tasks, reduces load on engineers, and improves response times, demonstrating practical value beyond theory. Governance, guardrails, and emergent behavior (Priority: 4/5): The discussion balances enthusiasm for emergent multi-agent intelligence with concerns about divergence, safety, and interpretability, arguing that cognition engines/guardrails are essential for enterprise adoption.

Key Arguments: AI progress is not only about scaling models vertically; it also requires horizontal scaling through collaboration among specialized agents. Human cooperation historically depended on language and culture; AI needs analogous protocols to enable shared intent and collective problem-solving. Current networking protocols move deterministic data, but agentic AI exchanges probabilistic cognition state, requiring new protocol layers. Enterprises need least-privilege, task-specific access for agents because agents are non-deterministic and can change behavior rapidly. Discovery and identity are the control points for trust, reputation, security, and commerce; they should not be monopolized by a single platform. Open, interoperable standards and decentralized identity/directory options are necessary for a healthy agent ecosystem. Multi-agent systems can already automate meaningful enterprise work, but shared context, negotiation, and governance remain human-mediated today. Guardrails are not optional in enterprise AI; they are a prerequisite for adoption and must evolve alongside model capability. Diversity among agents and models is valuable for resilience, redundancy, and emergent innovation, similar to best practices in distributed software systems.

Data Points: CAPE agent count: 20 agents - Cisco’s internal Community AI Platform Engineer system uses roughly 20 agents to manage cloud/SRE workflows. Automated task share: 40% - Pandey says CAPE has fully automated about 40% of tasks handled by the SRE team. Team load reduction: 30% - CAPE reduced load on the SRE team by about 30%. Response time improvement: hours to instantaneously - Automating SRE workflows with CAPE reduced response times from hours to near-instant. Tool calls: 100+ - CAPE performs more than 100 tool calls across cloud providers and cloud-native environments. Workflows covered: 10+ - The system tackles more than 10 workflows across observability, orchestration, networking, and security. Protocol layers proposed: 2 - Pandey proposes two new networking layers for AI: syntactic communication (layer 8) and semantic/cognitive communication (layer 9). Cognition protocol types: 3 - He describes semantic state transfer, latent space transfer, and compressed state transfer protocols. Human language step-function: ~70,000 years ago - Pandey cites the invention of language as the major step change in human intelligence evolution. Human-prep benchmark: 16 hours - Referenced in discussion of model evaluation: tasks around this duration are where human parallelization begins to matter more. Open-source community members: Adobe, AWS, Cisco, Nike, and others - The cloud-native operating excellence community around CAPE includes multiple enterprise participants. Demo agents: 4 agents - A healthcare demo shows four agents collaborating across scheduling, insurance, diagnostics, and pharmacy.

Pivotal Quotes: "What’s missing, Vijay says, is the internet of cognition." — Nathan Lambert: Opening framing of the episode’s central thesis about agent-to-agent collaboration. "We need to be able to reason about things, and we need to be able to persist around certain kinds of state... we are exchanging cognition state." — Vijay Pandey: Explaining why AI communication needs new protocol layers beyond the traditional OSI stack. "The collective is always greater than the individual." — Vijay Pandey: Describing the rationale for horizontal scaling through multi-agent collaboration.

Implications: If this vision holds, AI will shift from isolated super-models to distributed agent ecosystems with shared protocols, trust, and governance. That could make enterprise AI safer, more interoperable, and more powerful—while reducing concentration of capability and control.