Your website may need an interface nobody can see

Websites have spent decades learning how to present information and actions to people.

Now a second kind of visitor is arriving.

AI agents do not browse websites like people.

They can inspect a page, read its content and try to operate its controls. But asking a machine to infer an action from a visual interface is still asking it to guess.

Being readable is not the same as being usable.

A help page can explain how to complete a task. That does not mean an agent can reliably complete it.

A person can interpret a vague label, notice an exception and stop when something looks wrong. An agent needs the action, inputs, rules, consequences and permission boundary to be explicit.

One useful model for the emerging agentic web is that websites need to become:

• Readable
  The information can be accessed and understood.
• Discoverable
  The right route can be found.
• Callable
  The site can expose a defined action rather than making an agent imitate clicks.
• Payable
  Where money is involved, the transaction can happen through an explicit, permissioned route.

A second interface beneath the first.

WebMCP is a draft web specification that allows web applications to provide JavaScript-based tools to AI agents.

Instead of an agent studying a page and guessing which button means what, a website could declare the actions it supports and the information each action requires.

Chrome's early WebMCP material uses examples such as booking travel, filing a support ticket and navigating complex data. The proposal includes declarative tools for standard form actions and imperative tools for more complex interactions.

The visible website does not disappear. People still need clear language, accessible controls, context and judgement.

But beneath that interface, some websites may develop a second functional layer: less visual, more explicit and designed to be called by another system.

This is not just another technical markup job.

The difficult part is not exposing a function. It is deciding when the function is safe to use.

Chrome's own security guidance warns that WebMCP tools and their outputs must be treated as untrusted, that cross-origin interactions should be restricted, and that people should confirm consequential actions.

A callable website still needs authentication, review, consent, recovery and a clear account of what happened.

That is why agent experience is wider than whether a machine can find and press the right control. It includes whether the route is understandable, permissioned and recoverable.

Not every website needs to become a machine.

A publication may mainly need to be readable and discoverable. A service that lets people compare, book, buy, change or cancel something has a stronger reason to make its actions explicit.

The practical question is not “Does our website support WebMCP?”

It is: which useful actions should another system be able to understand—and what must remain visible to the person authorising them?

What BIC brings back.

Your website may eventually need two interfaces.

One that people can see, understand and trust.

Another that makes its available actions legible to machines.

A website is no longer only a place to visit. It may also become something another system can call.