Core Concept

Mental Model

The core insight of A2UI: agents express what they want the UI to do; the UI decides how to render it.

The Problem with Traditional Approaches

Raw Text Parsing

Agent outputs unstructured text. Client must parse, guess intent, and hope formatting is consistent. Brittle and error-prone.

Rigid JSON

Agent outputs custom JSON schemas per use case. No interoperability between different agents or UIs. Every integration is bespoke.

The A2UI Approach

Standardized Intents

Agent emits UI intents from a known vocabulary. Any conforming renderer can display them while maintaining visual flexibility. The protocol defines semantics, not presentation.

Separation of Concerns

Layer	Responsibility	Owner
Reasoning	Decide what to show/ask	Agent
Intent	Express UI needs structurally	A2UI Protocol
Rendering	Visual presentation	UI Framework

Example Flow

Agent reasoning (internal)

"User asked about the weather. I should show current conditions."

Agent emits A2UI intent

{
  "type": "Card",
  "title": "Current Weather",
  "fields": [
    { "label": "Temperature", "value": "72°F" },
    { "label": "Conditions", "value": "Sunny" }
  ]
}

UI renders appropriately

Web:Card component with styled fields

CLI:Formatted text block

Voice:"The temperature is 72 degrees and sunny"

Key Benefits

Model-agnostic

Works with any LLM that can emit structured output

Renderer-independent

Same intent works across web, mobile, CLI, voice

Explainable

Intents are inspectable and debuggable

Evolvable

New primitives can be added without breaking existing ones