The Website Has to Earn the Visit

An expense report is not a place I want to visit. I want to provide the receipts, explain what they were for, resolve anything missing, and submit the expense. The login screen, navigation, and wizard are the route the software requires, not the reason I use it.

An agent could prepare that transaction from information I already supplied and return only the questions and decisions that need my attention. The provider would remain responsible for its rules and records without owning the entire interaction. That is the consequential interface change behind agents: not another chat box added to a website, but the website becoming an optional client of the service.

The Interface Moves To The User

In his 2015 book The Best Interface Is No Interface, Golden Krishna walks through thirteen steps for opening a car door with a phone app. Most involve finding and operating the app. His alternative is proximity-based entry: approach with the key and pull the handle. The task never required a screen. The app introduced one.

No UI starts by removing interaction the task never needed. Asking an agent to open the car would still be unnecessary when pulling the handle expresses the intent. On the web, the next question is not only whether a task needs an interface, but why it must use the provider’s interface.

A web application bundles information, operations, and presentation. Its portal exposes destinations before knowing why I arrived; its filters let me progressively describe what I want. The provider chooses how I narrow those possibilities into an action.

A user-directed agent can separate those responsibilities when the provider exposes data and operations through authorized machine access. The agent interprets intent and presents what returns. Clicking through the existing website automates the old route; exposing the capability independently removes the need for it.

Whether that agent runs locally or through a service matters less than whose preferences govern it. My interface and context can remain familiar while the services behind them change, without each provider having to build my preferred client.

Ink & Switch makes this distinction in its 2025 essay Malleable Software: sharing data and composing interfaces are separate problems. An application can let data out while still forcing users between isolated environments. The authors argue for tools people can adapt where they use them. An agent that builds another disconnected application has not solved that problem; it needs to bring existing capabilities into an environment I can shape.

One of their examples comes from Atul Gawande’s Why Doctors Hate Their Computers. His 2018 reporting describes both the benefits of shared medical records and the burdens of rigid workflows. Neurosurgeon Neil Malhotra and software analyst Judy Thornton worked with their department to build an interface suited to neurosurgery visits while retaining the underlying system. That is the separation worth pursuing: preserve shared records and necessary rules while letting practitioners shape the interaction.

Forms Become Schemas, Not Destinations

A form exposes the shape of a transaction: required information, allowed values, and conditions for submission. An agent can translate my receipts into that structure, then ask about an ambiguous charge or missing project. Instead of discovering the transaction one field at a time, I review a prepared expense.

Tax preparation raises the stakes. The goal is a correct return, not a completed interview. An agent should assemble records and reduce transcription, while leaving uncertain tax treatment and missing facts visible for resolution. Removing data entry does not remove judgment.

The provider’s form can disappear while its requirements remain in a machine-readable contract. Authorization, validation, and policy still apply. A locally rendered review may even look like a form, but it appears to resolve uncertainty or confirm an action, not to make me manually reconstruct the request.

Don Norman’s The Design of Everyday Things distinguishes the gulfs of execution and evaluation: figuring out how to act and what happened. Agents can reduce the first while widening the second if they hide their decisions. I still need to understand, correct, and evaluate the action. That is where the remaining interface earns its place.

The exit path remains part of the feature. Stopping pending work differs from reversing a completed action. The service must report whether nothing happened, a submission was accepted, or a separate cancellation is required.

A Dashboard Guesses Your Question

A dashboard chooses variables, aggregation, time range, and emphasis before its viewer arrives. That is valuable for recurring, shared questions: an operations team needs stable definitions and alert thresholds. For an unexpected investigation, though, I must manipulate those choices to reconstruct the comparison I wanted.

The question should come first: where did latency change after this release, and what happened to retries? My agent can retrieve the observations and prepare the comparison. I should not have to choose between a table, time series, or topology before asking, or accept the first representation if it does not explain the result.

Google Research’s 2025 work on generative UI demonstrates interactive pages, tools, and simulations generated in response to prompts. It supplies a way to construct the view after the question is known, rather than placing every answer inside a fixed chat layout. It does not settle who controls the data or the agent.

The evidence must remain inspectable: sources, filters, units, transformations, and uncertainty. A presentation fitted to my understanding must not quietly fit the data to the conclusion I wanted.

The Storefront Becomes A Choice

Shopping makes the ownership question concrete. I want dresses in my size that can arrive before Friday. Retailers supply images, measurements, prices, inventory, and terms; my agent compares candidates through my preferred interface. The request survives across stores without rebuilding each search.

Approval must bind to the item, price, and delivery terms. If Friday delivery becomes unavailable, the old approval should not silently authorize a different purchase. The provider must return revised terms for review, not let the agent’s attractive rendering turn stale inventory into a promise.

The harder obstacle may be commercial. Storefronts promote products, encourage bundles, and control comparisons. Direct access weakens that control over attention. APIs already separate services from clients; agents make it easier to assemble an interaction without writing a custom client for every task. They do not guarantee providers will permit it.

Nor is the agent automatically neutral. Replacing retailer merchandising with an agent platform’s undisclosed ranking incentives would move the gatekeeper rather than give the user control.

An Experience Is Different

When I launch a video game, I choose another creator’s rules, pacing, and world. Those things are the product. An interactive story or authored visual explanation can offer the same reason to visit: its composition is part of its meaning, not an obstacle between me and a transaction.

WebGL, Three.js, and 3D assets have been available for years. The barrier was also the effort of designing and producing a coherent scene. I recently started using Blender with agent assistance to create assets beyond my familiar CSS tools. A first matching set made unrelated blog posts look interchangeable, so I replaced it with scenes specific to each story. Producing geometry became easier; deciding what it should communicate still required direction. The illustration here is part of that exploration.

An expense report does not need a Tron interface. It needs to become a reliable capability. Richer rendering belongs where the presentation contributes meaning, not where it makes an obligatory workflow more elaborate.

Text Already Travels

Text already shows that authored meaning can travel without its original renderer. An essay can preserve its words, sequence, and references in reader mode, a feed reader, or a screen reader. Typography and layout affect reading, and some figures or equations depend on arrangement, but a blog rarely needs its navigation rail to communicate the author’s thought.

An infographic’s composition may be the argument itself. It can remain a shared authored object, while a personal investigation calls for a different explanation generated around one question. The distinction is whether I came for the author’s interpretation or need a representation suited to my own task.

The Website Has To Earn The Visit

Provider interfaces can remain valuable as shared operational views, trusted confirmations, and chosen experiences. Users still need accessible controls and direct manipulation where a gesture is more precise than a sentence. None of that requires every service to own the whole interaction.

The useful split is between services agents can consume and experiences people choose to enter. A retailer can offer both purchasing capabilities and an editorial storefront worth visiting for its taste. What changes is the obligation to use the storefront. Reliable information and actions should not require accepting the provider’s entire theory of presentation.

If you own the data, let me ask for it. If you own the action, let my agent request it. If you want me to enter your presentation, make it worth seeing.