AI-native website interfaces

Evolving proposal · 26 July 2026

The better interface is not a chatbot added to a conventional page. It is a website whose underlying work is directly readable by people and AI agents.

Question

What should a website become when people and AI agents are both first-class readers?

Core idea

Current answer

Make the durable unit an AI-readable investigation object. The human interface is one view of that object—not the object itself.

The object should expose claims, evidence, relationships, confidence, work state, and possible actions directly. An agent can then inspect or extend the work without reverse-engineering a screen.

Why it matters

Evidence · working observations, not benchmark results

A conventional web interface hides meaning behind interaction choreography. An agent must translate intent into menus, tabs, clicks, and playback steps, then translate the resulting screen back into state.

intent menu tab click playback extract

  • Menus and tabs delay access to relationships until the right control is found.
  • Click sequences encode procedure, but rarely expose why each step matters.
  • Video-only evidence adds seek, playback, and transcription before claims can be compared.

How the parts relate

Map / relationships

  1. INPUT Question + evidence The material being investigated and the provenance behind it.
  2. DURABLE CORE Investigation object Claims · relations · confidence · work state · actions
  3. VIEWS Human page + agent context Different interfaces reading the same source of meaning.
A chatbot may be one view, but it is not the architecture.

One source of meaning can support a calm reading page for people, compact structured context for agents, and other interfaces later.

Work state

Verified
This proposal exposes its six parts as named semantic sections, with state labels that do not depend on color.
Uncertain
The shared schema, permission model, and best balance between machine detail and human calm still need testing.

A practical next step

Next actions

  1. Define a small public schema for the investigation object.
  2. Apply it to one real investigation, with claim-level provenance.
  3. Compare clicks, tool calls, time, and recovery against the current page.
  4. Revise this answer when the evidence changes.