Cisco's Outshift research arm has open sourced AGNTCY, a multi agent infrastructure project now hosted at the Linux Foundation with 65+ supporting companies including Dell, Google Cloud, Oracle, and Red Hat.
Cisco built its networking empire by owning the protocols — not the devices. MPLS, BGP, Spanning Tree: Cisco wrote the rules that competitors had to follow, then sold the switches that ran them. It's a playbook that's worked for decades. Now the company's research arm, Outshift, is trying to run the same play in agent infrastructure, and unlike most "AI vision" announcements, this one has shipped code, a governance structure, and 65 supporting companies.
The project is called AGNTCY (pronounced "agency"). Cisco open sourced it in March 2025, then donated it to the Linux Foundation on July 29, 2025, with Dell Technologies, Google Cloud, Oracle, and Red Hat as formative members. As of this writing it has more than 65 supporting companies. According to the Linux Foundation announcement, "Building the foundational infrastructure for the Internet of Agents requires community ownership, not vendor control," said Vijoy Pandey, general manager and senior vice president of Outshift by Cisco. "The Linux Foundation ensures this critical infrastructure remains open, neutral, and community-driven." That framing — community ownership over vendor capture — is the same argument the Linux Foundation has made for every major infrastructure project it hosts, and it's the same argument that made Kubernetes a standard rather than a product.
AGNTCY's technical steering committee includes seven organizations: Dell, Red Hat, Cisco (with two seats), Oracle, and Google Cloud. The project covers four problem areas that anyone who's tried to wire together agents from different frameworks will recognize immediately: discovery (how agents find each other), identity (cryptographic credentials that survive cross-organizational boundaries), messaging (via a protocol called SLIM — Secure Low Latency Interactive Messaging), and observability (debugging a workflow that spans three vendors' agents). It's interoperable with Anthropic's Model Context Protocol and the Agent2Agent protocol that was contributed to the Linux Foundation separately.
Internally at Cisco, the system is called CAIPE — the Community AI Platform Engineer. It runs 20 agents in production managing Cisco's cloud computing environments, according to Pandey on the Cognitive Revolution podcast. The CAIPE GitHub repository shows six named sub-agents: ArgoCD for continuous deployment, PagerDuty for incident management, GitHub for version control, Jira and Confluence for project management, Slack and Webex for team communication, and Kubernetes for cluster operations. The gap between "20 agents" and "six named sub-agents" on the public docs isn't a contradiction — Cisco runs 20 internally; the public reference implementation shows six. The architecture supports both multi-node deployment via A2A and single-node via MCP over stdio.
The 20-agent internal system that coordinates all of this is called JARVIS. It uses a LangGraph architecture with supervised agents, specialized agents, and reflection agents working in feedback loops — 15 or more specialized sub-agents, 40 tool integrations, 10 automated workflows, according to Network World. JARVIS is being refactored toward the AGNTCY architecture, which means Cisco is walking its own talk: the internal system and the open-source project share a dependency graph.
The Internet of Cognition whitepaper — the philosophical wrapper around the infrastructure — proposes three protocol classes for agent state transfer: Latent State Transfer Protocols for high-fidelity, low-latency communication; Compressed State Transfer Protocols to reduce payload sizes; and Semantic State Transfer Protocols for semantic understanding between heterogeneous systems. Whether those three classes become real standards or remain a taxonomy in a whitepaper depends on whether AGNTCY adoption spreads beyond the founding members.
The governance structure is where Cisco's play becomes legible. Open-source projects with vendor-led governance tend to reflect the founder's interests first and the ecosystem's interests second. AGNTCY under Linux Foundation governance is different in form, though not yet in practice — the TSC includes Dell, Red Hat, Cisco, Oracle, and Google Cloud, and any of them can veto changes that favor Cisco specifically. That neutralizes the obvious objection ("this is just Cisco trying to own agent infra") without eliminating the economic reality: Cisco wrote most of the early code, its employees maintain the key repos, and its internal systems are the reference implementation.
This matters beyond the enterprise networking crowd. The broader story is a platform war being fought on two fronts simultaneously. On one side, infrastructure vendors — Cisco, Dell, Red Hat, Oracle, Google Cloud — are building the open substrate through foundations, betting that a neutral governance layer wins the ecosystem's trust the way Linux won server infrastructure. On the other side, platform companies are acquiring agent startups and the engineers who built them.
Meta Platforms acqui-hired Dreamer on March 23, bringing in David Singleton (formerly CTO of Stripe and vice president of engineering in Google's Android unit), Hugo Barra (formerly vice president of product management for Google's Android operating system, then Oculus at Meta, then cofounder at Xiaomi), and Nicholas Jitkoff (formerly a design lead at Figma). They join Meta Superintelligence Labs under Chief AI Officer Alexandr Wang. Singleton confirmed the deal in a post on X, saying he had demonstrated Dreamer's technology to Meta CEO Mark Zuckerberg earlier this year. "It was clear right away that we share the same vision of the future: one where billions of people have the power to create software that makes their lives better," Singleton wrote. The Dreamer team will remain a standalone legal entity; Meta gets a non-exclusive license to the technology.
Singleton and Barra both built platform businesses at scale — Android's app distribution model, Stripe's API ecosystem. They're not joining Meta to ship another model. They're joining because the platform play is agent creation and distribution, and Meta has distribution at a scale no other company can match. The parallel to Cisco's strategy is precise in structure if not in intent: own the governance layer, and the applications built on top will follow.
Whether AGNTCY becomes the MPLS of agents or ends up as a well-engineered repo that nobody productionizes depends on one variable that's impossible to measure from the outside: whether enterprises beyond Cisco actually deploy it. The 65 supporting companies include enterprise buyers, but "supporting" and "running in production" are different things. The live demo — a multi-agent healthcare system spanning diagnostics, insurance, pharmacy, and scheduling, which Pandey walked through on the podcast — is impressive as a proof of concept and inconclusive as evidence of adoption.
The Cisco play, if it lands, is that the infrastructure layer becomes a utility and Cisco becomes the vendor-neutral authority that maintains it. That worked for networking because the protocols were genuinely hard to reinvent and the switching hardware had real margins. Agent infrastructure may not follow the same logic: the protocols are young, the compute is cheap, and the hyperscalers have strong incentives to define their own agent standards rather than cede that territory to a foundation Cisco helps govern.
What's worth watching: whether the AGNTCY TSC publishes a public roadmap with committed delivery dates, or whether it stays a Cisco-flavored vision until the founding members' economic interests and the community's needs diverge. The governance documents are live on GitHub. The code is live. The next question is who actually shows up to maintain it when the conference buzz fades.