The Web Isn't Forking. It's Being Hollowed Out.

In a single 72-hour window this month, Coinbase launched crypto wallets designed not for people but for AI agents, Cloudflare started automatically converting web pages to markdown whenever an AI system requested them, and OpenAI shipped tools that let agents install software dependencies, run scripts, and write files inside hosted containers. None of these companies coordinated their announcements. They didn't need to. They're all building for the same future.

The popular framing for what's happening is "the web is forking" — splitting into a human web and an agent web, the way a river divides around an island. Both channels flow. Both matter. Everyone gets served. It's a lovely metaphor. It's also dangerously wrong.

The agent web doesn't run parallel to the human web. It runs through it, extracting the valuable bits and discarding the rest. Your hero banner, your scroll animations, your carefully A/B-tested "Add to Cart" button placement — none of it registers with software that wants a JSON payload with price, availability, and a payment endpoint. That's not a fork. That's a hollowing out. And if you build, sell, or optimise ecommerce experiences for a living, the next few years are about to get very uncomfortable.

The Comfortable Lie of the Mobile Analogy

Everyone reaches for the mobile comparison. In 2007, the iPhone launched. The web already existed but it was built for desktops. What followed was a decade of rebuilding: responsive design, mobile-first frameworks, app stores, push notifications, GPS-aware services, tap to pay. The infrastructure was the same. The interfaces were fundamentally different. The mobile fork created Uber, Instagram, WhatsApp, and Snap — businesses that couldn't have existed on the desktop web because it lacked the primitives mobile clients needed: real-time location, always-on connectivity, camera-first interaction.

The agent web, we're told, will follow the same playbook. New primitives, new businesses, new trillion-pound companies. Just swap "mobile" for "agent" and the pattern repeats.

Except it doesn't. Because the mobile web had one crucial thing in common with the desktop web: a human being was still on the other end. A human who had preferences, emotions, brand loyalty, impulse-buying tendencies, and a finite attention span that could be captured and monetised. Mobile didn't eliminate the consumer. It gave the consumer a different screen.

The agent web eliminates the consumer from the transaction entirely.

When an AI agent purchases compute on demand, rebalances a DeFi portfolio, or buys raw materials for a manufacturing workflow, there is no consumer in the traditional sense. There is software executing against criteria. It doesn't care about your brand story. It can't be upsold. It won't abandon its cart because you asked for an email address. It has no cart. This distinction matters enormously, and almost nobody in ecommerce is reckoning with it.

Your Product Page Is Talking to Nobody

Here's a number that should make every agency and ecommerce team pause: Cloudflare serves roughly 20% of the web. When Cloudflare decided AI agents are first-class citizens of the internet — not bots to be blocked, but clients to be served in their preferred format — they made an infrastructure-level commitment to a world where software reads websites as routinely as humans do.

And software doesn't read your product page. Not the way you designed it.

A human lands on a product page and processes lifestyle photography, social proof, trust badges, urgency timers, cross-sell recommendations, and a carefully structured information hierarchy designed to guide them toward conversion. That page represents thousands of pounds in agency fees, months of UX research, and countless rounds of stakeholder feedback.

An agent hits the same URL and gets a markdown conversion of whatever structured data happens to be present. If you're lucky, it extracts the product title, price, availability, and maybe a description. If you're not lucky — if your product data lives in JavaScript-rendered components, behind lazy-loaded elements, or scattered across six different micro-frontends — the agent gets garbage. And unlike a frustrated human who might click around trying to find what they need, the agent simply moves on to a competitor whose data is machine-readable.

Cloudflare didn't stop at markdown conversion, either. As Search Engine Land noted, the implications alarmed the SEO community — and with good reason. They launched three companion features in the same release. An agents.txt file — the robots.txt of the agent web — letting sites declare what agents can and can't do. An AI Index where sites can make their content discoverable directly through Cloudflare's MCP server, bypassing Google entirely. And built-in x402 monetisation support, so site owners can charge agents for content access using the same payment protocol as Coinbase's wallets.

They're not building features. They're building an economic layer for a web where agents pay to access content. That's not a tweak to the existing model. That's a different model.

The ecommerce industry has spent two decades optimising for human psychology. Heat maps, eye-tracking studies, conversion rate optimisation, persuasion architecture. All of it assumes a human brain on the other end, processing visual information and making emotionally influenced decisions. When the buyer is software, that entire discipline doesn't become less important. It becomes irrelevant.

Stripe saw this before most. As VentureBeat reported, the era of the limited agent is waning — and Stripe's response was telling. Their Radar fraud detection system — built over years of machine learning calibrated on human shopping behaviour, things like mouse movement variability, browsing time, session patterns, device fingerprinting — had to be retrained from scratch. Every signal they'd built their fraud models on became useless when the buyer doesn't move a mouse, doesn't browse, and doesn't exhibit the behavioural variability that distinguishes a legitimate shopper from a bot. Because the buyer is, by any prior definition, a bot. Stripe had to build an entirely new fraud model for a class of customer that their entire system was previously designed to reject.

If Stripe threw out decades of fraud detection logic to accommodate agent buyers, what makes you think your conversion funnel is going to survive intact?

The Quiet Death of the Conversion Funnel

The conversion funnel — awareness, consideration, decision, purchase — has been the foundational model of digital commerce since the late 1990s. Every marketing strategy, every analytics platform, every ecommerce build is structured around the assumption that a customer moves through stages, and your job is to guide them through those stages as efficiently as possible. Reduce friction here. Add social proof there. Retarget the ones who drop off. The entire industry runs on this logic.

Agent commerce doesn't have stages.

An agent receives an instruction: "Purchase 500 units of SKU-7742 at the lowest available price with delivery by Friday." It queries available suppliers, compares pricing and logistics, validates payment credentials, and executes. The entire process takes seconds. There is no awareness stage because the agent already knows what it wants. There is no consideration stage because comparison happens computationally, not emotionally. There is no decision stage because the decision was made by the human who set the agent's parameters — possibly days earlier, possibly as a standing policy that runs indefinitely.

The infrastructure for funnel-less commerce is already in production. Google launched its agent payments protocol in September 2025. PayPal and OpenAI partnered on instant checkout in ChatGPT. Visa built a Trusted Agent Protocol and debuted it at NRF 2026 in January. Google announced the Universal Commerce Protocol — an open standard for agent-to-commerce interaction — and Stripe's Agentic Commerce Suite automatically supports it. Merchants who integrated Stripe's agent tools are already compatible with Google's agent shopping infrastructure without writing a single additional line of code. Brands including Urban Outfitters, Etsy, Coach, Kate Spade, and Revolve are onboarding now.

That's not a handful of start-ups running experiments. That's the dominant payment infrastructure of the western world aligning around a commerce model that has no funnel, no checkout flow, and no product page.

For B2B commerce, the implications are even starker. A procurement agent doesn't need to be nurtured through an email drip. It doesn't respond to case studies or whitepapers. It doesn't attend webinars. It reads your API, checks your data feed, evaluates your terms programmatically, and either transacts or moves on. The entire B2B marketing apparatus — content marketing, lead scoring, SDRs, demo calls — exists to compensate for the fact that human buyers are slow, emotional, and easily distracted. Agent buyers are none of those things.

The UX-Industrial Complex Meets Its Irrelevance

There's a multi-billion-pound industry built on making websites pretty and functional for humans. Design agencies, CRO consultancies, UX research firms, accessibility auditors, A/B testing platforms, personalisation engines, customer data platforms. This entire ecosystem exists because the interface between a business and its customers is a visual one, mediated by screens that human eyes look at.

The agent web doesn't have screens.

I want to be clear: human-facing web commerce won't vanish overnight. People will still browse, still impulse-buy, still enjoy the experience of shopping online. The human web will persist in the same way physical retail persisted after ecommerce — diminished in relative importance, consolidated around experiences that justify the format, but still present.

But the growth is going to be in agent commerce. Within 24 hours of Coinbase launching its agentic wallet infrastructure, ten thousand new AI agents registered wallets on Ethereum. That's not developer experimentation. That's an ecosystem forming in real time. Polymarket reported in early February that autonomous AI agents are now trading on their platform "in an attempt to subsidise their token costs" — agents trying to earn money to pay for their own compute. The economic loop is closing faster than the governance frameworks can keep up.

For agencies and ecommerce teams, the uncomfortable question is: what do you actually sell when the customer is software?

You don't sell design. Software doesn't see it. You don't sell persuasive copy. Software reads it for data and ignores the persuasion. You don't sell brand experience. Software doesn't experience brands. You don't sell conversion rate optimisation. Software doesn't convert through a funnel.

What you might sell — what the market is going to pay for — is structured data architecture. API design. Product information management. Schema markup. Machine-readable commerce protocols. The boring, unsexy plumbing that nobody wanted to invest in when the job was making things look nice for humans. The agencies that survive the agent web won't be the ones with the best Figma files. They'll be the ones who can make a product catalogue perfectly legible to software, with clean structured data, reliable APIs, and machine-readable policies on pricing, availability, returns, and shipping.

That's not a pivot. That's a completely different business.

Trust Is the Only Moat — and It's Vanishingly Thin

There's a fundamental tension at the heart of the agent web that nobody has resolved: the infrastructure assumes full autonomy, but nobody trusts agents enough to grant it.

Every serious security implementation treats the agent as a potential adversary, not a trusted employee. IonClaw — a Rust-based re-implementation of OpenClaw by Near.ai co-founder Ilya Polosukhin — sandboxes every single tool into isolated WebAssembly environments, on the assumption that any tool an agent touches is a potential compromise vector. OpenAI's shell tool includes organisation-level network allowlists and container isolation, on the assumption that agents will run untrusted code and the environment must contain the blast radius. Coinbase's agentic wallets use enclave isolation for private keys with programmable spending guardrails, on the assumption that the agent itself cannot be fully trusted with the assets it manages.

Notice the pattern: every serious player in the stack treats the agent the way a prison treats an inmate. Contained. Monitored. Given the minimum access necessary to perform a specific task. This is the correct mental model for 2026, and it's one that the TikTok tutorial crowd — the people claiming to turn fifty quid into three thousand pounds with an OpenClaw trading bot — have not remotely internalised.

One developer who built and tested an autonomous Polymarket trading agent reported that running the bot for two days racked up $200 in API costs alone, before generating a single dollar of return. Another found that Cloudflare blocks API requests from data centre IPs, requiring custom bypass infrastructure just to place orders. The gap between "agent that can trade" and "agent that trades profitably" is filled with infrastructure costs, security requirements, and operational complexity that no tutorial covers.

But the underlying premise — that agents can participate in economic activity and generate revenue — isn't a scam. It's the direction that Coinbase, Stripe, Google, PayPal, Visa, and OpenAI are all aggressively building toward, simultaneously, with billions in infrastructure investment. The question isn't whether agents will transact autonomously. The question is whether guardrails can be built fast enough to prevent very predictable disasters.

And every security incident — wiped databases, runaway message bots, credential leaks from unsupervised OpenClaw instances — pushes the trust timeline backwards even as the capability timeline accelerates. The businesses that solve trust, through transparency, audit trails, and demonstrated reliability, will have the only moat that matters in the agent economy. Because trust is earned slowly and destroyed instantly, and the agent web has been destroying it at a remarkable clip.

What Comes After the Hollow

The web isn't splitting into two equal branches. It's developing a new dominant layer — the agent layer — that will increasingly mediate how commerce happens, while the human layer becomes the shopfront rather than the engine room.

A developer who goes by ChatCut demonstrated this week what the emergent agent web actually looks like in practice. He connected OpenClaw to CogVideoX, then sent the agent an Amazon product link. The agent crawled the page, extracted product information and images, identified suitable assets for video generation, fed them into the model, and produced a user-generated-content-style product video — the kind brands pay creators a thousand pounds to produce. No human touched any step between "paste this link" and "here's your video."

That Amazon page wasn't designed for agents. CogVideoX wasn't designed to receive input from web crawlers. ChatCut wasn't designed as an orchestration layer. But because each service exposes its capability through APIs and structured data, the agent stitched them together into a workflow that no individual company planned. This is the pattern the infrastructure convergence makes inevitable: when content is available as markdown, search returns structured data, execution happens in containers, and payment flows through tokenised protocols, the agent doesn't need anyone to build an integration. It reads both services, understands both, and chains them on the fly.

For ecommerce businesses, the hollowing means three things.

First, your data is your product. Not your website, not your brand, not your checkout experience — your structured, machine-readable product data. If an agent can't find your inventory, read your prices, understand your shipping policies, and transact through your payment rails, you don't exist in the agent economy. Full stop.

Second, the metrics change. Conversion rate optimisation assumes human visitors moving through a visual funnel. Agent commerce metrics look more like API performance monitoring: response time, data accuracy, uptime, error rates. The businesses that win will be the ones with the cleanest, fastest, most reliable commerce APIs — not the prettiest storefronts.

Third, the value chain inverts. For 25 years, ecommerce has been a design-forward discipline. The highest-status, highest-billing work was the front-end: visual design, user experience, brand storytelling. The back-end — product data management, API architecture, structured content — was unglamorous plumbing. In the agent web, the plumbing is the product. And the plumber becomes the most important person in the building.

The mobile web fork created Uber, Instagram, and WhatsApp — companies built for humans holding different devices. The agent web won't create new apps for humans to use. It'll create new infrastructure for software to transact through. And the companies that emerge from it will be the ones that recognised, early enough, that making the web beautiful for humans was always a temporary competitive advantage.

One that expires the moment your most valuable customer doesn't have eyes.