Episode Summary
Executive Summary: Vijoy Ponde of Cisco Outshift argues that enterprise AI is moving from deterministic software to a probabilistic “internet of agents,” where specialized agents from different vendors collaborate like human SMEs. Cisco’s Agency open source project and Slim transport layer aim to provide discovery, identity, composition, deployment, observability, security, and real-time multi-agent communication.
Main Topics: Cisco Outshift’s role and investment thesis (Priority: 5/5): Vijoy explains Outshift as Cisco’s internal incubator focused on adjacent markets with technology and market risk, including agents and quantum-safe systems. Internet of Agents as the collaboration layer (Priority: 5/5): Cisco’s core thesis is that future business problems will be solved by ensembles of specialized agents that need a shared platform for discovery, identity, communication, and trust. Agency open source project and practical enterprise use cases (Priority: 5/5): Agency is positioned as the open source manifestation of the Internet of Agents, illustrated through multi-vendor enterprise workflows such as sales pipeline setup and SRE automation. Deterministic to probabilistic computing shift (Priority: 5/5): The conversation frames agentic computing as a higher abstraction layer than cloud computing, introducing both productivity gains and new operational pain around reliability, governance, and accuracy. Protocols, semantics, and transport layers (Priority: 4/5): The discussion distinguishes syntax-level protocols like A2A, MCP, and ACP from deeper semantic interoperability and from Slim, Cisco’s transport layer for secure real-time multi-agent messaging. Enterprise deployment modes for agents (Priority: 4/5): Vijoy outlines three maturity stages: deterministic agent graphs, semantically routed graphs with human steering, and eventual self-forming agent workflows. Evaluation, observability, and quantum-safe security (Priority: 4/5): Beyond communication, the platform must measure agent accuracy, manage distributed behavior, and ensure secure, quantum-safe operation across multimodal, stateful workflows.
Key Arguments: Agents are the next abstraction layer after cloud computing, replacing human SMEs and much of enterprise toil with specialized software collaborators. The Internet of Agents is needed because disparate vendor agents must discover, trust, and coordinate with one another in enterprise workflows. Existing SaaS integration patterns relied on deterministic interfaces and glue code; agentic systems replace them with probabilistic interfaces that need new governance. Enterprise adoption will not be fully autonomous at first; organizations will likely start with deterministic graphs, then add semantic routing, and only later reach self-forming agent systems. Protocol choice alone is insufficient: agent systems also need identity, discovery, deployment, observability, evaluation, and transport guarantees. Semantic interoperability matters as much as syntax; the industry is focusing on A2A/MCP/ACP mechanics while underestimating meaning-level alignment. Slim is intended to bridge agentic communication to infrastructure realities, providing security, real-time behavior, and multimodal state handling on top of existing transports. Cisco’s work demonstrates measurable ROI by removing support toil and accelerating common cloud-native tasks from days to minutes.
Data Points: Support engineers replaced: 3 engineers - Jarvis now handles a support desk that previously required three SRE/support engineers Task time reduction: from a day or more to minutes - About 30% of engineer tasks such as cluster setup, EC2 creation, identity setup, and LLM provisioning Query response time: seconds - Simple engineer queries, such as checking running Docker images, are reduced to seconds Jarvis internal agents: 20 agents - Jarvis is described as a multi-agent system with 20 agents behind one interface User interfaces for Jarvis: 5+ interfaces - Access via WebEx chat, Backstage, VS Code plugin, Jira, and CLI Accuracy expectation example: 90% confidence - Used to illustrate how probabilistic outputs from agents may differ semantically across vendors Current observed accuracy example: 60% accuracy - Illustrative contrast showing how stitched agent workflows may underperform per-agent claims Per-agent accuracy example: 95% accuracy - Illustrative claim that individual agents may each appear highly accurate, yet ensemble performance can drop Cisco scale projection: 10x to 100x larger - Vijoy says the Jarvis model at Cisco scale will be substantially larger than Outshift scale Deployment maturation stages: 3 modes - Deterministic sandwich, semantically routed graphs, and self-forming agents
Pivotal Quotes: "You replace SMEs with agents. That's the world where we're heading into where a bunch of these subject matter expert agents come together and solve for a particular problem." — Vijoy Ponde: Describing the central thesis of the Internet of Agents "To us, the Internet of Agents is that platform. It's an open, interoperable, quantum safe platform for agent-to-agent collaboration." — Vijoy Ponde: Defining Cisco’s platform vision for multi-agent coordination "We are moving from deterministic computing to probabilistic computing or a combination of deterministic and probabilistic computing." — Vijoy Ponde: Explaining the step-function shift underlying enterprise agent adoption
Implications: Enterprises should prepare for agent ecosystems, not isolated copilots. The winners will provide trust, identity, observability, semantics, and secure transport—not just model access or glue code.