Let me define the term first, because "Web4" gets used loosely.
By Web4 I mean this: AI agents running on decentralized infrastructure, with verifiable identities and auditable actions, so that no single company, platform or government controls them or can quietly override them.
Web2 gave us platforms. Web3 gave us ownership. Web4 is what happens when the users of the internet are increasingly not people but agents acting for people. And it raises a question we haven't answered yet: when an agent shows up to buy, book, negotiate or pay, how do you know who it is, who it works for, and whether it can be trusted?
The problem is already here
This isn't a 2030 problem. Banks are already living it.
In BioCatch's 2026 survey of more than 1,400 fraud and risk leaders across 25 countries, 80% said their institution had already encountered attacks using agentic AI. More telling, 72% worried it will be very difficult to tell legitimate AI-assisted actions apart from malicious ones once agents routinely start transactions.
Read that second number again. The issue isn't that agents are dangerous. It's that legitimate and malicious agents look the same. Today we solve trust with logins, cookies, device fingerprints and a human clicking "confirm". None of that works when the actor is software that never sleeps, runs in thousands of copies and has no face.
The obvious fix, and why it's the wrong one
The easy answer is to let the big platforms vouch for agents. Your agent is trusted because it runs on a major cloud, carries its badge, and that company can switch it off.
That works until you notice what it creates: a handful of companies deciding which agents may exist, what they may do and whose interests they serve. We'd be rebuilding the walled gardens of Web2, except this time the gardens contain the software that spends our money and speaks on our behalf.
That's the "superpower" problem. Whoever controls agent identity controls the agent economy.
What a trust layer no one owns looks like
The alternative is a shared, neutral registry, closer to how DNS or email work than to an app store. Every agent gets a portable identity, a public track record and a way to prove it did what it claimed, all recorded somewhere no single party can edit.
This is being built now:
- Identity and reputation. ERC-8004, the "Trustless Agents" standard, was created in August 2025 by authors from MetaMask, the Ethereum Foundation, Google and Coinbase. It defines three on-chain registries: identity (who the agent is), reputation (how it has performed) and validation (proof it followed its rules). It went live on Ethereum mainnet in January 2026.
- Payments. Coinbase's x402 protocol lets agents pay for services over standard web requests using stablecoins. Industry reports put it past 100 million transactions.
- Authorization. Google's Agent Payments Protocol (AP2), developed with more than 60 organizations, uses cryptographically signed "mandates" so an agent can prove what a human actually approved before it spends.
Put together, an agent can find another agent, check its identity and history, pay it, and leave a permanent rating, without a platform sitting in the middle as gatekeeper.
Let's be honest about where this is
I'm convinced of the direction, not naive about the timing.
- ERC-8004 is still a draft standard under review.
- Registrations are growing fast, but actual agent-to-agent spending is still tiny. The identity layer is running well ahead of the money.
- Regulation hasn't caught up. Who is liable when a verified agent still does damage? Today, nobody has a clean answer.
- On-chain doesn't mean unhackable. Smart contracts and bridges have been exploited for billions. A shared ledger makes tampering visible; it doesn't make attacks impossible.
So this isn't "blockchain fixes AI". It's narrower and more useful: agents need an identity and an audit trail that outlive any single company, and decentralized registries are the best design we have for that today.
Why I care, as someone from advertising
I've spent 18 years in ad tech, and our industry has already lived through a smaller version of this.
Programmatic advertising was built on machines buying from machines. It scaled beautifully, and then it filled with bots, spoofed domains and fraud, because nobody could verify who was on the other end of a bid. We spent a decade bolting verification on afterwards.
Agents are about to do the same thing to commerce, and at a much bigger scale. Very soon an agent will choose which brand to recommend, which offer to accept and which ad to show its human, if any. Brands will need to know they're dealing with a real agent acting for a real customer. Publishers will need to know an agent's spend is authorized. Consumers will need to know their agent isn't quietly being paid by the highest bidder.
That trust has to be designed in from day one this time, not patched on later. And it shouldn't belong to whoever happens to own the biggest model.
What to do now
If you're building or marketing in this space, three practical moves:
- Treat agents as a customer segment. Start asking how an agent, not a person, would discover, evaluate and buy from you.
- Watch the standards, not the hype. Identity (ERC-8004), payments (x402, AP2) and coordination (A2A, MCP) are where the real rules are being written.
- Design for verifiability. Whatever you build, make it easy to prove what happened, who authorized it and who benefited.
The agent economy is coming either way. The open question is whether its trust layer gets owned by a few, or shared by all of us. I know which one I'm betting on.
— Aditya Tibrewal
Sources
- BioCatch 2026 survey via Technology Decisions: https://www.technologydecisions.com.au/content/security/news/banking-security-leaders-see-ai-as-the-industry-s-top-threat-1370163798
- ERC-8004 authorship and draft status: https://stablecoininsider.org/erc-8004/
- ERC-8004 mainnet launch (Jan 2026): https://www.cobo.com/post/erc-8004-on-chain-identity-standard-for-ai-agents-the-future-of-agentic-wallets
- x402 transaction volume and AP2 mandates: https://www.livebitcoinnews.com/can-crypto-standards-make-ai-agents-truly-autonomous
- AP2, x402, A2A, MCP overview: https://erc-8004.quicknode.com/learn/agent-protocols-overview