Pillar
Web4: the web that acts on your behalf
Web4 is the working name for what happens when software stops waiting for clicks. Not a standard, not a product - a set of changes that are already visible in how services are built, addressed and paid for.
A definition worth defending
Web4 is the phase of the internet in which the primary user of a service is software acting on behalf of a person or an organisation, rather than a person operating an interface directly. Everything else follows from that sentence: if the visitor is an agent, the page layout matters less than the machine-readable description of what the service does; the login matters less than a verifiable claim about who the agent works for; and the checkout matters less than a payment method that works without a browser session.
That is a narrow definition on purpose. The term Web4 has also been used for the metaverse, for a "symbiotic web" of brain-computer interfaces, and by the European Commission for a strategy paper covering virtual worlds. Those usages exist and are not wrong, but they describe different things. On this site Web4 means the agentic web, and nothing else.
There is no W3C recommendation called Web4, no version number, no launch date. Anyone who tells you Web4 "starts" on a particular day is selling something. What exists is a cluster of shipping technologies that fit together suspiciously well, and a useful label for talking about them at once.
Web3 versus Web4, in one paragraph
Web3 was an answer to a question about ownership: who holds the asset, who controls the account, who can transfer value without asking a platform for permission. Web4 is an answer to a question about agency: which piece of software may act, on whose authority, and within what limits. The two are not competing narratives. Web4 consumes Web3 - the ownership and settlement primitives built over the last decade are precisely the ones an autonomous agent needs the moment it has to pay for something. The full comparison goes through it layer by layer.
The five building blocks
Strip out the vocabulary and Web4 rests on five components. Four of them work today. The fifth is where the interesting problems are.
- Agents that can plan and use tools A model in a loop with external systems, a memory and a goal. This is the part that arrived first and is now routine - see AI agents for how the different flavours (autonomous, multi-agent, browser, computer use) actually differ.
- Interfaces built for machines, not eyes A service that wants agent traffic has to describe itself in a way a model can consume: structured capability descriptions, stable endpoints, predictable errors. Tool protocols standardised this in 2024–2025 and it is now the least controversial layer of the stack.
- Identity for non-human actors An agent needs an identifier that cannot be forged and a credential proving whose authority it carries. Decentralised identifiers and verifiable credentials solve the cryptography; adoption is the bottleneck. See digital identity for AI agents.
- Payments that work at machine scale and machine speed Sub-cent amounts, thousands of times a day, without a card form or a human confirmation step. Stablecoin rails and HTTP-native payment schemes both do this now - the open question is not transport but authorisation and refunds. More on agent payments.
- A record of what happened that survives a dispute When an agent commits you to something, someone will eventually want to know what it was told, what it was allowed to do, and what it actually did. This is the layer where AI and blockchain infrastructure genuinely meet - and the layer furthest from being solved.
Machine-to-machine, and why it changes the economics
The web's business model assumes a human on the other end. Ads are priced against attention. Free tiers are subsidised by conversion. Rate limits assume a person cannot click faster than a person. Every one of those assumptions breaks when the visitor is software that never sleeps, never sees an ad, and reads a hundred competing offers before deciding.
The likely replacement is metered, per-request pricing paid directly by the calling agent - which is exactly why machine payments are a load-bearing part of this story rather than a crypto-flavoured detour. A service that can charge a tenth of a cent per structured answer does not need an ad network, a signup funnel or a session cookie. It needs an identity check and a payment header.
If agents become the main readers of the web, attention stops being the currency and execution starts.
Autonomous services
Once services are addressable by machines and payable by machines, they can be composed by machines. A booking is not a form any more; it is a negotiation between a buyer's agent and a seller's agent over availability, price and terms. That is the shape of autonomous commerce, and pieces of it are already in production - discovery and qualification much further along than settlement and dispute handling.
The same logic applies to the interfaces humans still use. A phone call handled end to end by a real-time voice agent is a Web4 service wearing a Web1 interface: the caller experiences a conversation, while behind it a model is querying calendars, checking stock and writing records through exactly the machine interfaces described above.
What Web4 is not
- Not a blockchain project. Ledgers appear where trust between strangers is required - identity, settlement, provenance. Most of the stack does not touch one.
- Not AGI. Nothing here requires a model that is smarter than a person, only one reliable enough to be trusted with a bounded task and a spending limit.
- Not the metaverse. Different term, different decade, occasionally the same word.
- Not a replacement for the human web. Agents are an additional class of visitor, not a substitute for the existing one. The web will serve both for a long time.
What would have to be true
It is worth naming the conditions under which this does not happen. If liability for autonomous action stays unallocated, enterprises will keep a human approval step in every loop, and "autonomous" becomes a marketing adjective. If identity standards fragment into per-platform API keys, agents stay locked inside single vendors and cross-organisational commerce never materialises. If payment authorisation cannot express a meaningful spending mandate, consumers will simply not hand over the keys.
None of those is a technical problem. All three are governance problems wearing technical clothes - which is a reasonable summary of where this whole field currently stands.
Frequently asked
What is Web4 in simple terms?
Web4 is the stage of the internet where software acts on your behalf instead of showing you a page to click through. You state a goal, an AI agent works out the steps, calls the services it needs, and reports back. The web still exists - you just stop being the one operating it.
Is Web4 the same as the agentic web?
Yes, as the term is used here. "Agentic web" is the more descriptive name and "Web4" is the shorter one. Both refer to an internet whose main consumers are autonomous agents rather than human readers.
Does Web4 replace Web3?
No. Web4 builds on it. The ownership, settlement and identity primitives developed under the Web3 label are the ones agents need in order to transact and to prove what they were authorised to do. Web3 answered who owns something; Web4 asks who may act on it.
When will Web4 arrive?
It is not a release, so there is no date. Tool use and voice agents are already in production. Agent identity and machine payments are in the standards-and-pilots phase. Liability and dispute handling for autonomous action are unresolved, and that is what will set the actual pace.