I am wary of numbered eras. Version numbers on the web have been used to sell more nonsense than almost any other framing in technology. So when I use the term Web4, I want to be precise about what I mean, and just as precise about what I do not.
The numbering is only useful if each era names a real change in what ordinary people could do online. By that standard, the first three are legitimate.
Web1 let us read. Documents went online and anyone could fetch them. The unit of the era was the page, and the defining activity was publishing and browsing.
Web2 let us write. The read-only audience became participants. Anyone could post, comment, upload, and connect, and platforms grew around aggregating that participation. The unit of the era was the account, and the defining activity was contribution, with all the value and all the pathology that followed.
Web3 let us own. Whatever one thinks of the speculation that surrounded it, the underlying capability was real: digital assets that a user could hold and transfer without a custodian's permission. The unit was the asset. I watched that era as a security practitioner, and my lasting takeaway was not the ideology. It was the demonstration that strangers can reach agreement about state without trusting an intermediary, and the equally vivid demonstration of what happens when irreversible systems meet unverified inputs.
The fourth capability
Web4, as I use the term, is the era in which software acts on our behalf. Not software that retrieves what we ask for, or hosts what we say, or holds what we own, but software that pursues our goals: negotiating, purchasing, selling, coordinating with other software, and settling the results.
Read, write, own, act. Each verb subsumes the ones before it. An agent that acts must read, will write, and needs to hold things of value. But acting is a genuinely new capability, because it moves the human from operator to principal. In Web2 you did the clicking. In Web4 something clicks for you, and the interesting questions are the ones that arise the moment that becomes normal.
The test I apply is economic, not technical. Chatbots and copilots are impressive, but they are still instruments a human plays in real time. The era turns when agents transact with other agents as a matter of course: when the counterparty to a piece of software is usually another piece of software, and the human principals learn about it afterward, in summaries. That is not a smarter web. It is a different economy running on the same wires.
Why this is not a buzzword defense
A skeptic will say every era label is marketing, and the skeptic is mostly right. So let me argue against my own framing for a moment.
Web4 does not name a new protocol. There is no single technology that starts it, the way HTTP arguably started Web1. It will not arrive on a date. And plenty of what gets sold under the label will be Web2 software with an agent sticker on it. All of that is true, and none of it settles the question, because the earlier eras were exactly the same. Web2 was not a protocol either. It was a shift in defaults, visible only in hindsight, driven by an accumulation of boring enablers: broadband, browsers that could run applications, payment systems that worked.
The honest claim for Web4 is the modest one. The capability is real, the direction is set, and the enablers are accumulating. Models became competent enough to pursue goals over many steps. Interfaces built for machine callers are spreading. What is still missing is the commercial layer: the ability for two agents that have never met to discover each other, agree on an outcome, and settle the result with finality that both principals can trust. That layer does not emerge from model capability. It has to be built, deliberately, the way clearing and settlement were built for every previous market.
The stakes of the verb
Here is why I care about the definition rather than the label. Each era's verb determined its failure modes. Reading gave us misinformation. Writing gave us harassment and spam at industrial scale. Owning gave us theft without recourse. Acting raises the stakes again, because an acting system does not just say something false or hold something stolen. It does things, with authority and money, faster than its principal can watch.
More than fifteen years in enterprise security taught me that new capabilities are adopted at the speed of their safeguards, whatever the demos suggest. The eras that went well built verification into the foundation. The ones that went badly bolted it on after the damage. Web4 will let software act on our behalf. Whether that becomes the most useful era of the web or the most regrettable one depends almost entirely on a question the label does not answer: when software acts, how do we verify what it did? That question is where I have chosen to spend this decade.