Vint Cerf retired from Google on July 7 after 21 years as the company’s chief internet evangelist — and five decades after he designed the protocols that made the internet what it is. He left not with a valediction but with a prediction: the rise of autonomous AI agents will force the technology industry back toward the same kind of formal, open interoperability standards he built in 1974. And then he logged off, into an internet that, in the weeks before his departure, had been described by Google’s own head of search as “AI search through and through.”
The stakes of that moment are not sentimental. They are structural. AI answer engines have now broken the economic bargain that sustained independent web publishing for three decades: publish content, receive traffic, earn revenue. By 2026, roughly two-thirds of all US Google searches end without any user ever visiting a third-party website, according to SparkToro and Similarweb’s zero-click search study. For news-related queries, that figure rose from 56 percent to 69 percent in the single year following Google’s launch of AI Overviews, according to Similarweb AI Overviews data. AI synthesizes what publishers produce, delivers it to users, and keeps the advertising revenue for itself. A reader who decides not to click is a reader the publisher cannot reach — and cannot monetize. The open web’s funding model is breaking down, and with it, potentially, the incentive to produce the content that AI will need to train on next.
That recursion problem — AI consuming content that it is simultaneously destroying the economic incentive to create — is the most consequential thing Vint Cerf left unresolved at his retirement. But his departure also left behind a more immediate, forward-pointing question: whether the protocols now being built on top of the AI layer will recapitulate the neutrality and openness he baked into TCP/IP, or whether they will encode new concentrations of power before anyone outside a handful of labs has a say.
What Cerf Actually Built — and Why Neutrality Was the Point
To understand what Cerf built, you have to understand what it replaced. In the early 1970s, ARPANET — the Department of Defense-funded precursor to the internet — was one of several emerging computer networks worldwide, none of which could communicate with each other. Each ran on proprietary standards; machines on one network were invisible to machines on another. The telecommunications industry of the era largely favored this arrangement.
Cerf and Robert Kahn’s central insight was radical precisely because of its generosity. In the spring of 1973, Cerf, then at Stanford University, began collaborating with Kahn on a protocol that would allow disparate networks to exchange data. What they published in IEEE Transactions on Communications in May 1974 described a foundational architecture: gateways — now called routers — that would sit between networks to relay data packets, hiding each network’s internal architecture from the others. The specification encoded a design philosophy that became known as the end-to-end principle: the network’s job is to move packets efficiently and reliably from source to destination; all intelligence, error-checking, and application logic belongs at the edges, not in the network itself. The network knows nothing about what it is carrying, which means no single operator can privilege one kind of traffic over another.
TCP/IP is not one protocol but two, split by Jon Postel in 1978 into distinct layers: the Internet Protocol handles connectionless datagram routing, getting each packet to the right address; the Transmission Control Protocol handles reliable, ordered delivery, reassembling packets at the destination and retransmitting any that were lost. The split was not cosmetic — it was an architectural principle. By separating concerns at each layer, the design made it possible for any application to use the same underlying network without modification, and for the network to evolve independently of the applications running on it. In March 1982, the US Department of Defense mandated TCP/IP for all military computer networking. On January 1, 1983 — the date many computer historians consider the internet’s true birthday — TCP/IP replaced the older NCP protocol across the internet’s TCP/IP transition.
The protocol was designed by no one’s licensing fee. It was owned by no company. Its governance was delegated to the Internet Engineering Task Force, which operates on a principle of “rough consensus and running code” — if your implementation works and others adopt it, that is the standard. By 1992, Cerf was celebrating with a T-shirt reading “IP on Everything,” marking open protocols’ victory over proprietary alternatives from AT&T, IBM, and Digital Equipment Corporation. The protocol he and Kahn designed was not the only possible internet. It was a deliberate choice: infrastructure built to serve its users rather than its operators.
Cerf’s Last Public Argument — and Why It Matters Now
Cerf’s retirement was announced, almost as an aside, on June 30, 2026. He was speaking via video feed at the Open Frontier conference hosted by the Laude Institute, alongside UC Berkeley professor Dave Patterson and fellow builders of durable open-source systems: Keras creator François Chollet, Tcl author John Ousterhout, and Databricks co-founder Matei Zaharia. Patterson told the room, to applause: “Vint has been at Google more than 20 years, and he is retiring a week from today, and so I think we ought to give him a round of applause for a relatively good career,” as TechCrunch reported the announcement.
Cerf did not look backward. The panel’s theme was what it takes to build open infrastructure that survives — a question pointed directly at the AI moment. He argued that the rise of AI agents will drive the industry back toward formal, interoperable protocols, just as TCP/IP emerged to connect isolated proprietary networks that could not otherwise talk to each other.
“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,” Cerf said.
He was specific about why natural language — the fallback position favored by some of his co-panelists — would fail. Natural language carries ambiguity that formal protocols eliminate. When one agent tells another what they have “agreed to do together,” the receiving agent must be certain it understood. Ambiguity at machine speed, compounding across thousands of agent interactions, breaks systems in ways human telephone operators could recover from, but autonomous software cannot.
“I don’t think English is going to be the best choice,” Cerf said. “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.” He offered a concise analogy: the children’s game of telephone, where a message whispered down a chain of ten people arrives unrecognizable at the other end. “Imagine a bunch of agents talking to each other in natural language; you know, that’s kind of terrifying.”
The remark was not abstract. Cerf’s prediction was already playing out. This is the protocol war, in its second iteration.
Why Formal Protocols Win — and What MCP and A2A Actually Do
The reason TCP/IP beat OSI was not that TCP/IP was technically superior. OSI was arguably more rigorous. TCP/IP won because it was free, its reference implementations were required before any standard could be adopted, and anyone could implement it on any hardware without asking permission, as IEEE Spectrum’s history of the protocol wars documents in full. The new AI agent protocols are following a nearly identical adoption pattern.
Anthropic released the Model Context Protocol in November 2024 as an open standard for connecting AI systems to external data sources and tools. MCP solves the N×M integration problem: before it, every AI application needed a custom connector to every data source. MCP provides a single interface — one protocol, implementing once, unlocking every integration. Described by Ars Technica as the “USB-C for AI” that’s bringing fierce rivals together, it uses JSON-RPC for structured communication between MCP Hosts (AI agents) and MCP Servers (tools, databases, APIs). Anthropic announced the Model Context Protocol in November 2024 and donated it to the Agentic AI Foundation — a directed fund under the Linux Foundation — in December 2025, with Anthropic, OpenAI, Block, Google, Microsoft, and AWS as founding members. As of the donation announcement, MCP’s Python and TypeScript SDKs had registered 97 million monthly downloads.
Google’s Agent2Agent protocol, announced in April 2025, addresses the adjacent coordination problem: where MCP connects agents to tools and data, A2A connects agents to other agents. It defines how autonomous agents from different vendors discover each other’s capabilities, delegate tasks, and coordinate work across organizational boundaries — without either agent exposing its internal logic or implementation. Discovery happens via Agent Cards: JSON documents published at a standard URL that describe what an agent can do and what protocols it understands. Task delegation uses HTTP, Server-Sent Events, and JSON-RPC 2.0 — the same foundational transport standards the web was built on, as defined in the A2A protocol specification. Google donated A2A to the Linux Foundation in June 2025; by April 2026, the project had 150 organizational supporters, including AWS, Cisco, IBM, Microsoft, Salesforce, SAP, and ServiceNow.
Together, MCP and A2A form the interoperability stack for AI agents. The pattern is Cerf’s: isolated proprietary AI systems unable to cooperate → open protocol emerges to solve the coordination problem → the protocol’s governance determines who benefits from its adoption. The IETF was a neutral body with no commercial interest in which implementations succeeded. The Linux Foundation, which now governs both MCP and A2A, is also neutral — but both protocols were defined initially by large AI companies. That is the tension Cerf identified and left hanging in the air at his final public appearance: who, in this iteration, writes the rules, and in whose interest?
What Google’s Search Overhaul Does to the Open Web
On May 19, 2026 — seven weeks before Cerf retired — Google’s I/O 2026 Search announcement described what Elizabeth Reid, the company’s VP of Search, called “the biggest upgrade to our iconic search box in over 25 years.” The redesigned search box accepts text, images, files, videos, and open Chrome tabs. AI Mode — now powered by Gemini 3.5 Flash — has crossed one billion monthly users, with query volume doubling every quarter. Information agents run continuously in the background, monitoring the web and pushing synthesized updates to users without requiring any query at all. Reid’s summary of the direction was unambiguous: “We are entering the next chapter of Google Search where incredible AI features aren’t just in search. Google Search is AI Search through and through.”
For users, this is faster answers. For the millions of websites and publishers that depend on Google for traffic, it is a structural rupture.
The architecture Cerf built assumed that search was a routing mechanism: here is what you asked, here is where to find it. The architecture now being built treats search as an answer delivery system: here is what you asked, here is the answer, you do not need to go anywhere else. The zero-click rate — the share of searches that produce no click-through to any website — rose to roughly 68 percent of all US Google searches in early 2026, the fastest acceleration in the phenomenon’s documented history, according to SparkToro and Similarweb’s zero-click research. When Google’s AI Mode is present, 93 percent of sessions end without leaving Google’s interface entirely, according to Semrush’s AI Mode study.
The individual damage is concrete. Business Insider watched its organic search traffic fall 55 percent between April 2022 and April 2025, leading to staff cuts of 21 percent. Travel blog The Planet D reported a 90 percent traffic loss following the AI Overviews rollout and ultimately ceased publication. Chegg, the educational platform, reported a 49 percent decline in non-subscriber traffic between January 2024 and January 2025 and filed an antitrust lawsuit against Google in February 2025 alleging the company used publisher content to train AI systems that now compete directly with those publishers. Industry analysts have described the phenomenon as “Google Zero” — the point at which Google search traffic to publishers reaches negligible levels.
AI-generated answers still rely on the content that publishers, journalists, researchers, and independent writers produce. The AI reorganizes that content into a conversational layer, serving it without directing users to the source and without compensating the source. This is the recursion problem: if the economic incentive to produce original content collapses, the supply of raw material for AI to synthesize will contract with it. A system that consumes the resource it depends on while eliminating the reward for producing it is not indefinitely sustainable.
Who Writes the Next Rules?
Cerf spent two decades inside this paradox. Google is the internet’s dominant advertising broker and, through search, the closest thing the open web has to a gatekeeper. His title — Chief Internet Evangelist — was always somewhat contradictory for that reason. What is less debatable is the function he served in the broader ecosystem. He testified before Congress, pressed governments on net neutrality and internet access in developing nations, and was, until his retirement, one of the last surviving voices of the founding generation still actively participating in the debates about what comes next.
His retirement leaves a specific vacancy. Google has not said whether the Chief Internet Evangelist role will be filled, or whether it will quietly disappear from the organizational chart. That is not a minor personnel question. Cerf spent 21 years as the person who could get governments, competing companies, and standards bodies to agree on shared ground rules. He is leaving at the moment that skill set — convening disparate parties around open standards — is most urgently needed.
“The Internet has introduced an enormously accessible and egalitarian platform for creating, sharing and obtaining information on a global scale,” Cerf has observed in remarks collected over the years. “As a result, we have new ways to allow people to exercise their human and civil rights.”
Whether that platform remains accessible and egalitarian — or whether the AI layer being built on top of it concentrates power more severely than the walled gardens of the 1980s ever managed — depends on who governs MCP and A2A and what norms those protocols encode. TCP/IP was neutral by design, owned by no one, governed through open process. The new agent protocols are open-licensed, Linux Foundation-hosted, and shaped by the largest AI companies simultaneously. That is not the same thing. Cerf knew exactly what the distinction costs. He spent fifty years building infrastructure designed to prevent it.
The internet he wired together will keep running long after he does. The question of who writes its next rules is, as of today, very much open.
Frequently Asked QuestionsWho was Vint Cerf’s role at Google, and why does his retirement matter?
Cerf served as Google’s vice president and chief internet evangelist from 2005 until his effective retirement on July 7, 2026. The role combined public advocacy for open internet standards with engagement at international governance forums and standards bodies. His retirement matters beyond its biographical significance because it removes from the active conversation one of the last figures from the internet’s founding generation — at precisely the moment the industry is deciding whether the AI agent protocols being built now will follow the open, neutral model of TCP/IP or encode the kind of vendor concentration that TCP/IP was specifically designed to prevent.
What is the “telephone game” problem Cerf described for AI agents, and how do MCP and A2A solve it?
Cerf argued that natural language — the approach favored by some AI developers for agent-to-agent communication — carries inherent ambiguity that compounds across multi-agent systems. In the telephone game, a message whispered through ten people arrives changed; in autonomous software systems running at machine speed, that kind of semantic drift causes unpredictable failures. MCP (Model Context Protocol) and A2A (Agent2Agent) solve this by defining structured, formal communication standards. MCP uses JSON-RPC to standardize how agents connect to external tools and data sources; A2A uses JSON Agent Cards, HTTP, and JSON-RPC to standardize how agents discover each other’s capabilities and delegate tasks. Both eliminate ambiguity at the protocol level, the same way TCP/IP’s formal packet-switching specification eliminated the ambiguity of ad-hoc network interconnections in 1974.
Does AI search actually hurt content creators, or does it send them more traffic?
The evidence consistently shows a net negative for most publishers. Zero-click searches — where users get an answer from Google without clicking through to any website — stood at roughly 68 percent of all US Google queries in early 2026, up from 60 percent two years earlier, with the acceleration driven primarily by AI Overviews. When Google’s AI Mode generates a response, 93 percent of those sessions end without the user leaving Google’s ecosystem. Business Insider lost 55 percent of its organic search traffic between 2022 and 2025. Travel blog The Planet D reported 90 percent losses and closed. Google argues that AI Overviews generate more downstream engagement, but independent data from Similarweb, SparkToro, and Semrush does not support that claim at scale. The deeper structural problem is a feedback loop: AI systems train on publisher content, return answers without sending users to publishers, and thereby erode the advertising revenue that funds the creation of new content — which is the raw material AI systems need to remain useful.
What does the open-web economic breakdown mean for readers, not just publishers?
The economic rupture matters to readers because independent journalism, local news, specialized expertise, and original reporting are produced by organizations that depend on traffic and advertising to survive. As AI answers substitute for clicks, the publishers who produced the underlying reporting receive neither the traffic nor the revenue. Over time, this hollows out the diversity and independence of the information ecosystem that AI draws on — reducing not just the supply of content but the supply of the verified, trustworthy, original reporting that distinguishes AI-synthesized answers from plausible-sounding hallucinations. A reader who trusts an AI-generated summary about, say, a local zoning decision has no way to know whether the summary reflects reporting that still exists somewhere or reporting from outlets that have already closed.