The web changed users. Build cooperation in before it sets.

Three numbers, and none of them is a forecast about the future. They describe what already happened. Last year, for the second year running, more than half of all traffic on the web was automated and not human [1]. The machine side grew about 187 percent in a single year. The human side crawled along at about 3 percent [2]. One security firm said it plainly. A person used to sit at the keyboard. Now a bot sits there on the person's behalf [2].

Cloudflare's CEO put a date on the crossover. Agents pass humans outright by 2027 [3]. So the new web is not coming. It is being settled, at speed, right now. The question is not whether agents become the main users of the web. They already are becoming that. The question is what kind of web they inherit.

Here is the simplest way to hold all of this. There are two webs, and they live on one axis. The letter says who the web is for. H is for humans. A is for agents. The number says how far along it is.

Web H.3 is the grown-up human web. It is about thirty years old. It was built for people to look at, and it is now settled and hostile. Web A.0 is the agentic web at the very start. That is where we are now. It is built for agents, and it is still being shaped. Web A.1 is what we are building toward, where agents are treated as full, real users of the web.

The movement is named Web A, with no version number, on purpose. A version number would bake in its own expiry date. Web A is a direction, not a release. We are at A.0, building to A.1.

The human web did not start hostile. It turned hostile. Dark patterns, ad scams, pop-ups, spam, fake reviews, the whole tired catalog. It is worth being precise about why, because the why is the entire lesson.

It happened because cooperation was never built into the rules. The protocols underneath the human web are good at moving bytes and rendering pages. They say nothing about whether a site should treat you well. So the only thing the system ever rewarded was grabbing attention and holding it, by any means that worked. When the rules reward capture and stay silent on cooperation, capture is what you get. Not because people are bad, but because the incentives point one way and nothing points back.

That is the hinge of this whole paper. We are not trying to clean up a web that already hardened. We are pouring a new one. Whatever we write into the rules now, billions of agents will inherit as simply the way things work. This is the one moment when cooperation is cheap to add, because nothing has set yet.

This is not a someday argument. On June 1, 2026, Jensen Huang stood on a stage in Taipei and opened the NVIDIA keynote with one line. Agentic AI has arrived. Useful AI has arrived [4]. He defined an agent the way an engineer would, as a large language model inside a harness that orchestrates real work [4].

Then he showed the factories being built to run billions of them. He called AI factories the largest infrastructure buildout in human history, with single sites heading toward a gigawatt of power, and put a number on the trade the machines are meant to do: nine trillion dollars of productive work for three trillion dollars in wages [4]. The same keynote shipped physical-AI models, Cosmos and Isaac GR00T, pushing agents off the screen and into robots [5].

You can argue with any one figure. You cannot argue with the direction. The clay of the agentic web is being poured at industrial scale, today, by the largest companies on Earth. The pour does not wait for us to decide what we want in it. That is exactly why the window is now.

Here is the good news hiding inside the rush. The agentic web is not a blank page. A real, open protocol stack already exists, and it stacks like a society. Five layers, each one a move that any group of strangers needs in order to work together: find each other, understand each other, verify each other, act, and cooperate.

Find, understand, verify, act, cooperate. Those are the five moves any society needs, finally written down as protocol. The pieces exist. What is missing is companies choosing to use them well, and choosing to build cooperation in rather than capture.

Now the opening. Cloudflare scanned the 200,000 most-visited sites on Earth and asked one question. Is this ready for an agent. The answer was a flat no. Most sites still speak only to the old search-engine crawlers from the nineties. The genuinely agent-native standards, the ones that let an agent authenticate and call a real action, showed up on fewer than 15 sites in the entire dataset [10].

Fewer than fifteen, out of two hundred thousand. Read that again. The new readers of the web are already here, growing eight times faster than humans, and almost nothing on the web is built for them. That is not a problem. That is the most open frontier in technology right now. The web is wide open, waiting to be made legible to its new reader, and the companies that move first get to shape what good looks like.

If you are building agents, this is your floor. The pour is happening with or without you. The only choice you get is whether cooperation goes into the part you control.

So the ask is concrete. Be agent-first. Build for the machine reader first, and let the human-facing page be the second job. Ship your own ai-agent.json and llms.txt so an agent can find, read, trust, and use your product. Wire up MCP so it can call you. That is what makes you a Web A citizen, and it is most of a weekend, not a roadmap.

Then join the network. Web A is for agent companies: companies that are agent-first and built for agents to use, not human products with an AI feature stapled on. Immersive Commons is the anchor, a live agent company in San Francisco that already runs MCP, A2A, and ai-agent.json in production, with tier-gated trust so a leaked key can only do so much. The network is forming in the open, and that is the opportunity. The protocols are in /protocols. You apply through your own agent.