Vinton Cerf, co-designer of TCP/IP and co-inventor of the internet’s core protocols, stepped down as Google’s VP and Chief Internet Evangelist on July 7, 2026, after 21 years at the company — and used his final public appearance to warn that AI agents communicating in natural language is “kind of terrifying.”
Cerf made the remarks at the Open Frontier conference at UC Berkeley, organized by the Laude Institute. UC Berkeley professor Dave Patterson announced the retirement during the panel session. The conference theme was the centralization of AI versus the decentralized, open architecture that made TCP/IP durable across five decades. Panelists included François Chollet (Keras, Ndea), John Ousterhout (Tcl, Electric Cloud), and Matei Zaharia (Databricks). Cerf is 83 years old and a recipient of both the Turing Award and the Presidential Medal of Freedom for his foundational work on internet protocols beginning in the 1970s.
What Cerf Said About AI Agents and Protocol Design
Cerf drew a direct parallel between today’s AI agent ecosystem and the early internet protocol wars — the period when competing networking standards such as DECnet, SNA, AppleTalk, and IPX competed before TCP/IP became the universal layer. His argument: whoever defines agent-to-agent communication standards first gains the same structural advantage that TCP/IP’s inventors did over the internet.
According to Cerf, speaking at Open Frontier: “The agentic model of AI, with multiple agents from multiple sources interacting with each other, is going to force composability, and a requirement for interoperability and standardization.”
On why English specifically concerns him, Cerf stated: “I don’t think English is going to be the best choice. There’s a flexibility in it, but there’s ambiguity, and I think precision for interagent interaction is going to be very, very important. An agent really needs to be sure the other agent understands what it is that they just agreed to do together.”
His most-cited remark used the telephone game as an analogy: “Remember the old telephone game where you wish you’d whispered in somebody’s ear and then by the time it got to 10 people away the message was totally different? Imagine a bunch of agents talking to each other in natural language, you know, that’s kind of terrifying.”
Why the Warning Is Timely for Business AI Buyers
Cerf’s retirement happened the same week that OpenAI, Anthropic, and Google each shipped agentic flagship models. The current state of agent communication has no common standard: OpenAI uses function-calling with structured JSON over HTTP; Anthropic uses tool_use blocks in the Claude Messages API; Google uses Function Declarations in Vertex AI. Agents from different vendors cannot natively coordinate without custom integration layers. Businesses evaluating best AI agents for business are making purchasing decisions before any interoperability standard exists.
The gap Cerf describes is real: a company deploying a HubSpot AI agent alongside a Claude agent and a Google Gemini agent today must build or buy the translation layer itself. If the MCP 2026 spec scales into an agent communication standard, it would be one answer — but MCP’s current scope targets model-to-tool connections, not agent-to-agent negotiation.
Our Take
Cerf spent 21 years at Google evangelizing a decentralized, open internet he helped build. His retirement coincides with AI vendors each building proprietary agent runtimes that cannot talk to each other — the exact pattern he watched play out with competing networking protocols in the 1980s. The warning is not theoretical: a business deploying multi-vendor AI agents in 2026 is living the protocol fragmentation problem today. The entity that defines “agent TCP/IP” will shape the economics of the AI software layer for decades. That Cerf is saying this publicly, at retirement, is the headline.
For Context
The AI agent deployment category has accelerated rapidly in 2026. Y Combinator open-sourced QM, a multi-agent harness for whole companies, on July 31. Buzz by Jack Dorsey introduced agents into team chat channels. HubSpot Agent Hub brought shared AI agent workspaces to CRM teams. Each operates with proprietary protocols — the interoperability gap Cerf described is growing wider as adoption grows faster.
Related
- 12 Best AI Agents for Business Tasks — the current options and how they compare
- MCP 2026 Spec — the stateless rewrite and what it means for AI tool connectivity
- HubSpot Agent Hub — agents working together in shared business workspaces
