Home / Explainers

Explainers

Frontend is Dead

Anatomy of a paradigm shift in which agents redesign how humans interact with websites: by removing the need to interact with them.

Danilo Giudice · May 17, 2026

Type

Explainer: from zero, with numbers and sources

Time

11 min read

Level

Beginner to intermediate

What you’ll get

• Why the frontend is losing interface, persuasion, and decision
• The new stack of the agentic web: MCP, A2A, x402, ERC-8004
• The questions that tell you whether an agent can read a website

This article was first published in Italian on Medium on May 17, 2026.

Abstract

In The Agentic Economy I tried to describe the shift from a web built around UX to a web optimized for MX, Machine Experience: an infrastructure where autonomous agents become the main actors in search, comparison, execution, payment, and verification. The idea came from a trivial fact. I had changed the frontend of a web interface so that an agent could navigate it without getting stuck. It was a small technical tweak, but looking back, it had changed who the interface was implicitly meant for: the machine. The predictable consequence, at least to me, is a World Wide Web where the frontend declines and is replaced by protocols that only agents can query. Because the last time you searched for something on Google, you probably stopped at what Gemini generated.

Chapter I: Frontend is Dead

The web frontend has based both how it works and its success on three monopolies: interface, persuasion, decision. It’s losing all three.

For twenty years it was the main surface where humans met and interacted with a digital service. You opened a website, scrolled the page, read a visual hierarchy, let a funnel guide you, and clicked a CTA. Then you bought, booked, subscribed, or simply left. We see this sequence as ordinary and natural only because we grew up inside it, unaware that we were taking part in a user journey designed by someone else.

The frontend is a technical compromise that became a culture, then an industry, then a dogma.

But the agentic economy breaks every one of its paradigms (and builds new ones).

Soon each of us will have a personal agent with access to our context, preferences, cognitive limits, calendar, budget, and email. That agent will make decisions for us, and we will no longer have to walk through the persuasive surface a company designed to turn us into customers.

When we delegate all of this, websites won’t disappear. They will simply stop being the place where the decision happens.

Chapter II: The frontend as a perceptual regime

The frontend was the point where a digital service became readable, desirable, comparable, and buyable. It was the membrane between database and psychology, between product and attention, between infrastructure and behavior.

A homepage is not just a homepage.

It decides what comes first and what comes after, which problem deserves a name, which pain must be amplified, and which benefit must look inevitable. It also decides which information is visible, which is hidden, which is postponed, and which is buried under two clicks and a modal.

So for years the frontend was the persuasive theater of the attention economy.

A well-made interface reduces complexity, makes a system accessible, and protects the user from chaos. It gives shape to something that would otherwise stay a pile of data, endpoints, states, errors, and policies. For the same reason, companies have always spent a lot of money to make interfaces an extension of their product. They pay marketing experts, UX microcopy writers, and designers: dozens of professionals behind a single landing page.

But the frontend, especially on the commercial web, was also something else: the controlled environment where the user was led to a decision (hero section, “most popular” badge, pricing tiers designed to make you pick the middle one, reassuring microcopy, a CTA repeated every three scrolls, exit-intent pop-ups, countdown timers, hand-picked reviews).

We called all of this User eXperience.

But of course, it was just choice architecture, controlled by the vendor. The vendor decided how you should see the service and how you should reach your choice.

This was the sovereignty of the frontend. And this sovereignty is what’s dying.

Chapter III: Website will survive, but…

The phrase “frontend is dead” is obviously a provocation. A useful one, though, because it forces us to separate two things we often confuse:

1. the frontend as a technical artifact
2. and the frontend as the symbolic center of the digital experience

The first will survive. The second won’t.

Websites will keep existing. Some will be needed for branding, trust, storytelling, compliance, documentation, onboarding, public relations, aesthetics, and institutional memory. Some will become premium surfaces, editorial spaces, immersive environments, places to tell stories. Others will stay simply because the legacy world never dies: it piles up in layers, it rots, and someone in an IT department keeps it alive in their never-ending backlog.

Website will survive, but…

In many cases websites will become human fallbacks on top of machine-readable services.

Value will move to the substrate the human eye can’t see: APIs, MCP, CLIs, databases, knowledge graphs, agent-readable endpoints, reputation layers like EIP-8004, machine-readable policies, identity systems, machine-to-machine payments like x402, and so on.

The death of the frontend has already begun, and not because of autonomous agents. It starts, for example, when the web stops being a network of destinations and becomes a layer of answers. Google has been doing this for years, but AI Overviews made the process explicit. Search used to be a list of links. Now it’s a summary on top of the links.

According to Pew Research, which analyzed about 69,000 Google searches, users who encountered an AI summary rarely clicked on a link in the summary itself: it happened in just 1% of visits to pages with a summary. Another study, reported by the Guardian, claims that a site previously ranked first in a search result could lose about 79% of its traffic for that query if results were delivered below an AI Overview. Google disputes the methods behind these studies and how representative they are, but the signal still matters: when the answer is good enough, the click becomes optional.

When another medium produces the output with less friction, the user stops using the old one.

Chapter IV: Agents don’t browse

A few days ago I asked Cowork to evaluate the best flights for my partner’s business trip. It was the first time I relied on an agent instead of searching for flights myself across the various aggregators.

A note on method: it wasn’t a real agent in the strict sense. It was Claude Cowork, which isn’t trained for the job of personal “travel planner.”

I gave it a few non-negotiable criteria: convenient times, a short distance from the venue, a reasonable price, and as little friction as possible.

This is worth stressing, because the real world isn’t sorted by a single filter. And humans often avoid putting that much effort into minor tasks, like choosing a hotel with a window in the bathroom.

The real world is full of trade-offs. The cheapest flight can be the worst one. The airport that looks close can be a logistical hell. A convenient time can become inconvenient if you land too late (and the agent accounted for this too, without me saying a word).

A low price can hide baggage fees, transfers, dead time, stress, connections, and micro-costs. The comparison site shows you these separately precisely because its job isn’t to protect you. Its job is to convert you.

Read that sentence again.

The agent connected to Kiwi and Lastminute, gathered options, produced a report, picked the best combination, and booked it. I didn’t browse, open twenty tabs, or do that tedious work I hate. I didn’t even consider whether it was right or not. I simply trusted it, because the output looked very reliable.

I delegated, because otherwise what the fuck would agents be for?

Read that sentence again.

UX doesn’t disappear right away, but it gets bypassed. The decision can be built and made outside the website: by an agent that uses that website as its source of truth, and by me asking the agent questions about that source of truth, without ever entering the damn source of truth.

The agent talks directly to the protocols: APIs, MCP servers, databases, endpoints, inventory systems, payment rails, booking engines, policy layers. It turns the web from a space you browse into a space of capabilities.

And in an economy where the marginal cost of agentic search collapses, forcing humans to browse becomes an archaic form of inefficiency.

Chapter V: SaaS is dead, frontend is next

Here’s the paradox: while SaaS is dying, companies are using more software than before.

Classic SaaS comes from a template. It solves a recurring problem, builds an average product around it, and sells it to thousands of companies. Then those thousands of companies adapt to that shape.

With AI the paradigm flips: instead of distilling one product across thousands of companies, the product is conceived, grown, and standardized inside a single company. Call it AI Engineer, AI Architecture, or Forward-deployed Engineering: these are already systems with job postings on Indeed growing 729% year over year.

And clearly this upends SaaS, because it wipes out product standardization in favor of something that goes beyond custom: a product born from one company’s specific need, which serves its workflows and nobody else’s.

Chapter VI: The frontend as a CLI printed to personal specs

The website serves a standard representation, and the browser receives it. Then the user’s personal agent reads the DOM, CSS, content, metadata, network requests, structured data, API calls, terms, pricing, policies, and affordances.

Next, it checks whether that representation suits the human it’s protecting. If it doesn’t, the agent transforms it.

It becomes a personal cognitive layer. If a user is dyslexic, the agent has that information, because it lives in its personal harness: preferences, limits, habits, accessibility settings, declared memory, maybe inferred memory, local tools, context.

When the human opens a website, the agent increases the spacing, changes the font, simplifies blocks of text, reduces information density, breaks up paragraphs, removes pointless animations, highlights decision patterns, and maybe reads parts out loud.

If a user is color-blind, the agent corrects the palette, the contrast, color cues, charts, and calls to action that rely only on color.

I love dark themes. Always, no exceptions. The website has no dark mode? Doesn’t matter. The agent injects one.

Maybe a user is a developer and wants to skip the parade of marketing punchlines and go straight to the documentation.

The agent hides the theater, brings the substrate to the front, and turns the same website into four completely different realities.

At that point, does “the website” still exist? Or are there only personalized renderings of a service? Schrödinger would say the two states are in superposition until you open it (the website).

This is my favorite image: a CLI printed to your personal specs. It’s like a home 3D printer, but for digital things. Your agent takes capabilities, context, policies, protocols, data, and reputation, then prints the minimum interface you need to complete that action.

For some it will be a table, for others a chat, for others a dashboard, and for others nothing at all, just the outcome.

Chapter VII: Machine Experience vs User Experience

In the first article I described this shift as a move from the UX Web to the MX Web: a web where the primary unit is no longer the human session but an agentic pipeline made of discovery, quote, payment, execution, proof, and reputation update.

Here the point gets more radical.

Many of today’s paradigms (UX/UI, user journeys, funnel design, information architecture, and conversion design) will have to be rethought, because they assume a completely different audience.

So design will have to stop designing only for human eyes and start designing for systems that can read, filter, compare, verify, and negotiate.

Can an agent parse this page? Can the claims be verified (because a human might fall for them, but an agent won’t)? Can a machine read the pricing? Are the policies structured? Are the limits explicit? Are the capabilities declared?

Looking trustworthy is no longer enough. You have to be queryable.

And the most uncomfortable question for many companies will be this one: how do I get chosen by an agent designed to protect the user from my own company?

And that’s exactly why the question is interesting.

Chapter VIII: The new stack: MCP, A2A, x402, ERC-8004, and agent-readable manifests

The agentic web will need a new grammar. Yes, an agent can navigate a website built for human eyes, but that’s still an awkward, in-between phase.

The destination is a callable web.

Anthropic introduced MCP in November 2024 as a protocol for connecting AI systems to data sources, tools, and external applications. It reduces the need to write custom connectors for every combination of model and source. This way a model can interact with tools and context in a standard way.

If you like, this is already part of the death of the frontend, because MCP doesn’t force the agent to interpret a page.

A 2025 paper on MCP points out that reliance on textual interfaces limits foundation models’ real-world interaction. It describes MCP as the response to a proliferation of frameworks with distinct tool interfaces.

And then there’s A2A, Agent-to-Agent, the core of The Agentic Economy.

If MCP is about how an agent interacts with tools and data, A2A is about how different agents can discover each other, communicate, and coordinate. The technical literature on the A2A protocol already focuses on security, Agent Cards, authentication, task execution, and interoperability with MCP.

Then you need payments.

This is where protocols like x402 get interesting, because they bring back an old idea buried in HTTP: 402 Payment Required. They also fix another tedious thing we won’t miss: subscriptions, accounts, human checkouts, forgotten passwords, and credit card forms. All of that goes away in favor of granular payments for access to resources, APIs, data, compute, and capabilities.

Then you need trust.

And here come layers like ERC-8004, built around identity, reputation, and validation for trustless agents. A 2025 comparative paper analyzes A2A, AP2, ERC-8004, and other protocols as a design space for trust between agents. It distinguishes between claim, proof, stake, reputation, and constraint. The conclusion is intuitive but important: reputation alone is not enough, because agents inherit LLM-specific weaknesses like prompt injection, hallucination, deception, sycophancy, and nudge-susceptibility. We need hybrid architectures that are easier to verify.

So the commercial website of the future might need less of a homepage and more of a manifest.

Something like:

/agent.json
/capabilities
/pricing.json
/policies.json
/mcp
/proofs
/reputation
/terms-machine-readable
/security
/refund-policy
/sla
/data-portability

Here an ontological shift takes shape.

Chapter IX: The agent will be the best signal-to-noise separator ever built

Suppose the personal agent becomes a filter built to protect your intent. It strips out noise, marketing fluff, and dark patterns. It shows you only what counts, queries the web on your behalf, and leaves the decision to you only where it really matters.

Which dogmas will fall first?

SEO, GEO, attention span, user journeys, and hero sections will probably lose the status they have today. Like all things, though, they will make room for something else. Agentic discovery, maybe?!

But agentic discovery is a different story.

Then there’s another issue that, as an anarchist, I care about only up to a point. The legal one.

I don’t want to cover it here, and I can’t.

Chapter X: What should companies do?

I don’t think the answer is “stop investing in the frontend.” That would be stupid. But if this thesis is even partly true, then companies need to start designing for agents too.

The end.

I’ll repeat the rest of the week for you, parrot-fashion

I keep up with AI, so on Monday morning you know what to try and what to ignore, with sources you can check.

Your email

Email me on Mondays

Leave this field empty

One email every Monday morning, in Italian. Unsubscribe with one click. Privacy

Prefer Telegram? News and tutorials go up there, in Italian, as soon as they come out.

Notify me on Telegram