Some tech leaders have talked about how software should be rethought in the age of AI, and others have talked about how UI and UX would need to be rethought, but Sam Altman believes that the entire internet might need to be rethought in the AI era.
“Feels like a good time to seriously rethink how operating systems and user interfaces are designed,” Altman posted on X. “(also the internet; there should be a protocol that is equally usable by people and agents)” he added.
The post is deceptively simple, but the implications are sweeping. The internet as it exists today was designed for humans — people who click links, fill out forms, navigate menus, and interpret visual layouts. AI agents do none of these things naturally. They scrape, call APIs, and parse structured data, often fighting against a web that was never built with them in mind.
A Protocol Layer For The Agentic Era
Altman’s vision isn’t new to him. Speaking at a Sequoia Capital event, he had sketched out what such a future might look like — something akin to a new protocol on the level of HTTP, where components are federated, and agents are constantly exposing and consuming tools with authentication, payment, and data transfer capabilities all built in at a foundational level. “Everything can talk to everything,” as he put it then. The difference now is that the argument is being made more urgently and publicly, as AI agents move from experimental curiosity to active participants in real workflows.
This also ties into Altman’s broader ambition for OpenAI to become what he has called the “default interface to intelligence” — not just an app, but an operating-system-level platform. A world where agents are first-class citizens of the internet would require exactly the kind of infrastructure OpenAI is positioning itself to build or influence.
The Broader Chorus
Altman is not alone in sensing that a reckoning is coming. NVIDIA CEO Jensen Huang has argued that AI will become “the operating system of every single industry going forward,” and that enterprises will need to completely reinvent their computing stacks to keep up. Box CEO Aaron Levie has suggested that the final form factor of AI may not be chat at all, but autonomous agents running silently in the background — a paradigm that the current web is particularly ill-suited to support. Former Google CEO Eric Schmidt has gone even further, predicting that static user interfaces will disappear entirely, replaced by interfaces generated on the fly.
Andrej Karpathy, meanwhile, has been making the case from an architectural standpoint: as agents proliferate, people may need entirely new tools — something like an “agent command center” IDE — just to manage them. The design requirements of a multi-agent world, he argues, look nothing like what developers use today.
Why This Is Hard
The challenge Altman is pointing to is more fundamental than it might appear. HTTP, HTML, and the broader architecture of the web were designed to serve documents to humans. Even modern APIs, while more machine-friendly, are designed with human developers as the primary audience — documentation, rate limits, authentication flows, and billing are all human-oriented constructs. An agent-native internet would need to rethink all of this from first principles.
There’s also the question of trust and identity. If agents are acting on behalf of humans across the internet — making purchases, signing contracts, managing communications — the web needs mechanisms to verify who an agent is, who authorized it, and what it is permitted to do. Sam Altman’s separate World project is already pushing in this direction, launching AgentKit as a way for verified humans to grant AI agents limited credentials. But that is a small piece of a very large puzzle.
The Stakes
The company — or consortium — that ends up defining the protocol layer for an agent-friendly internet will have extraordinary leverage over how the next era of computing unfolds. Altman knows this. His call to “seriously rethink” operating systems, interfaces, and the internet itself is as much a strategic signal as it is a technical observation. OpenAI already has 800 million weekly active users and the resources to shape what comes next.
The question is whether this rethinking happens through open, collaborative protocol design — the way HTTP and TCP/IP emerged — or whether it gets captured by a handful of dominant platforms. For now, Altman seems to prefer the former, at least in principle. Whether the incentives of a $800 billion company allow for that kind of openness is a different matter entirely.