An agent really needs to be sure the other agent understands what it agreed to do
Open standards enabled internet ubiquity by allowing any entity to connect using common protocols—a model Cerf says AI agents will need as their numbers grow. Natural language alone won't suffice for reliable agent-to-agent communication; AI systems need precise, unambiguous protocols to ensure mutual understanding of commitments.
- Vinton Cerf, 83, co-invented the networking protocols underlying the modern internet
- Open standards allowed any entity to connect to the internet using common protocols
- Natural language lacks the precision required for reliable agent-to-agent communication
- Transformative technologies like Google, Amazon, and Netflix scaled by building on open internet infrastructure
Internet pioneer Vinton Cerf argues AI development should follow three key internet principles: open standards over closed systems, standardized agent communication protocols, and platform-based architectures enabling broader innovation.
Vinton Cerf, now in his eighties, helped build the protocols that made the modern internet possible. This week, speaking at the Open Frontiers conference alongside computer scientists and Databricks cofounder Matei Zaharia, he offered a warning and a roadmap for artificial intelligence: learn from what worked before, or risk repeating the mistakes of closed systems.
The parallel Cerf draws is striking. The internet succeeded, he argues, precisely because no single company owned it. When a university network in California needed to talk to a government research lab, and both needed to connect with a commercial internet service provider, they could do it. They all spoke the same technical language. The rules were open. If you could find someone to connect to and you followed the protocol, it worked. That simplicity of access, that refusal to lock the system behind proprietary walls, is what made the internet ubiquitous.
AI is approaching a similar inflection point, Cerf believes. As the number of AI agents multiplies—systems talking to systems, making decisions, coordinating actions—the demand for interoperability and standardization will become unavoidable. The question is whether the industry will embrace that demand or resist it. Cerf's answer is clear: embrace it, or watch the technology fragment into incompatible silos.
But there is a subtlety here that matters. When humans speak to an AI agent, we use natural language—English, Mandarin, Spanish, whatever our native tongue happens to be. That works fine for a person asking a question and getting an answer back. It breaks down, however, when two AI agents need to coordinate with each other. Natural language is ambiguous by design. We rely on context, on inference, on the fact that the same word can mean different things depending on who is speaking and what they are trying to accomplish. An agent cannot afford that luxury. When one system agrees to perform a task for another, both parties need absolute certainty about what has been committed to. "I don't think English is going to be the best choice," Cerf said. Precision matters more than familiarity. The protocols that allow agents to communicate with each other will need to eliminate vagueness, to create a shared understanding that leaves no room for misinterpretation.
The third principle Cerf emphasized is perhaps the broadest: transformative technologies do not remain stand-alone products. They become platforms. Google did not build the internet; it built a search engine on top of the internet. Amazon did not invent networking; it created a marketplace and then a cloud infrastructure that others could build on. Netflix did not create the protocols; it created a service that millions of developers could integrate with. The internet itself was the enabling layer. Millions of smaller companies, countless individual developers, built their fortunes on that foundation.
AI will follow the same trajectory, Cerf suggests, if the industry allows it to. The real impact will not come from a single AI system, no matter how powerful. It will come from AI becoming the foundation upon which others build—the thing that enables people to do things they want to do, things they could not do before. That requires resisting the temptation to lock the system down, to make it proprietary, to extract maximum value from control. It requires, instead, the same openness that made the internet work.
The stakes are high. The choices made now—about whether AI development embraces open standards or retreats into closed systems, about whether communication protocols are shared or hoarded, about whether the technology becomes a platform or remains a product—will shape what AI becomes and what it can do. Cerf has seen this movie before. He knows which ending he thinks is better.
Notable Quotes
In the case of internet, it only worked because it was going to be distributed to begin with. We left the rules very open.— Vinton Cerf
I don't think English is going to be the best choice. There is ambiguity, and precision for inter-agent interaction is going to be very, very important.— Vinton Cerf
The Hearth Conversation Another angle on the story
When Cerf talks about open standards, is he really arguing against companies like OpenAI or Anthropic making money from their models?
Not exactly. He's not saying companies can't profit. He's saying the underlying protocols—the way systems talk to each other—need to be open. You can build a proprietary service on top of open infrastructure. Google did. Amazon did. The internet itself was open, but that didn't stop anyone from getting rich.
But doesn't that require a kind of restraint that tech companies don't naturally have?
Yes. That's the whole point. The internet worked because the people designing it made a deliberate choice to leave the rules open. They could have locked it down. They didn't. Cerf is saying AI needs that same deliberate choice, and it's harder to make now because the incentives point the other way.
He mentions that natural language won't work for agent-to-agent communication. Why is that such a big deal?
Because it's the difference between a system that can be scaled and one that breaks under its own weight. If two AI agents are trying to coordinate and they're using English, with all its ambiguity, they'll misunderstand each other. At scale, that's catastrophic. You need a language—a protocol—that is mathematically precise.
So we're talking about inventing a new language for machines to speak to each other?
In a sense, yes. But it's not new. We've been doing this since the early days of the internet. TCP/IP is a language. HTTP is a language. They're not natural languages. They're designed to be unambiguous. Cerf is saying AI needs something similar.
And the platform idea—is he saying AI should be like electricity? A utility that everyone builds on?
That's the aspiration. The internet became a utility because it was open and standardized. AI could become the same. But only if the industry resists the urge to keep it proprietary. That's the hard part.