Skip to content
AX Framework

The Agent Experience Framework

The web was designed for humans. Now design for agents.

AI agents are becoming a new class of digital users. AX Framework introduces the principles, patterns, and tools for building websites and applications that agents can understand, navigate, and interact with reliably.

GitHub repository coming soonv0.1 · Draft proposal
Human

User interface

Pages & menusContent & layoutButtons & formsToasts & spinners
AI Agent

Agent Experience layer

Appli­cation

Capabilities

The agent learns what the product can do before it does anything.

agent → GET /ax.json

{
  "capabilities": [
    { "name": "search_products", "risk": "read" },
    { "name": "add_to_cart", "risk": "low" },
    { "name": "place_order", "risk": "high",
      "requiresApproval": true }
  ]
}
Humans reach an application through a user interface. AI agents reach the same application through an Agent Experience layer made of capabilities, context, actions, and feedback.

A new user has arrived

Same product. A different kind of user.

For thirty years, digital products have been designed around one assumption: a person is on the other side of the screen. Agents break that assumption. They don't see layout, they don't remember your redesign, and they need to be sure before they act.

Human experience · UXAgent experience · AX

A strikethrough price means “on sale” to a person. An agent needs the original price, the current price, and the currency as data.

northwind.example/p/trail-runner
capabilities: search_products, add_to_cart, start_checkout, start_return
product: { sku: "TRAIL-RUNNER-42",
  price: { amount: 129, was: 149, currency: "EUR" },
  delivery: { minDays: 2, maxDays: 4, businessDays: true } }
action: add_to_cart({ sku, quantity: 1..10 })
  idempotent: true · risk: low
result: { status: "completed", cart: "cart_8f2", version: 2 }
policy: returns { windowDays: 30, excluded: ["final_sale"] }
  approval: required above 100 EUR

What is Agent Experience?

A definition

AX proposal

Agent Experience (AX) is the practice of designing digital products and services so AI agents can discover capabilities, understand context, execute actions, recover from errors, and achieve goals reliably, safely, and efficiently.

AX complements what already exists

AX is not a new protocol and does not replace your interface, your API, or your accessibility work. It is a way of designing across them for a user that operates differently.

UX design

Already provides

Interfaces people can understand and use.

AX adds

The same intent, expressed for software acting on a person's behalf.

APIs

Already provides

Programmatic access designed for developers who read the docs first.

AX adds

Runtime discovery, risk labels, approvals, and receipts for agents that read nothing in advance.

Accessibility

Already provides

Semantic structure that assistive technology can interpret.

AX adds

Builds on that structure. Accessible products are already partway to agent-ready.

Web standards

Already provides

HTML, schema.org, OAuth, HTTP semantics, and protocols like MCP.

AX adds

A design discipline for using them together, around tasks and outcomes.

The AX Framework

Six pillars, one loop.

Every agent task follows the same arc: find out what is possible, understand it, act, and confirm the result. Trust bounds the whole loop. Recovery keeps it going when something breaks.

Framework overview
  1. 01DiscoverabilityCan an agent discover what a product can do?
  2. 02ComprehensionCan an agent understand content, context, and constraints?
  3. 03ActionabilityCan an agent execute supported actions?
  4. 04Feedback & ObservabilityCan an agent understand outcomes and state changes?
  5. 05Trust & ControlCan an agent operate within clear permissions and safety boundaries?
  6. 06Recovery & ResilienceCan an agent handle errors, ambiguity, and interruptions?

From UX to AX

Each era added a new way in.

Graphical interfaces made computers usable by people. APIs made products usable by other software. Conversational interfaces let people use language. Agentic experiences let software pursue goals on a person's behalf.

  1. 01

    GUI

    Graphical interfaces

    People see, point, and click. Meaning lives in layout and convention.

    Primary user
    People
    Unit of design
    Screens, controls
    Contract
    Visual conventions
  2. 02

    API

    Programmatic interfaces

    Developers read the docs once, then write code that calls the product directly.

    Primary user
    Developers' code
    Unit of design
    Endpoints, schemas
    Contract
    Documented interfaces
  3. 03

    Conversational UI

    Natural-language interfaces

    People describe what they want. The product interprets intent and responds in kind.

    Primary user
    People, via language
    Unit of design
    Messages, intents
    Contract
    Language understanding
  4. 04Emerging

    Agentic Experience

    Agent experience layer

    Software decides how to reach a goal on someone's behalf, at runtime, within limits.

    Primary user
    Agents acting for people
    Unit of design
    Capabilities, actions, receipts
    Contract
    Declared, verifiable contracts

Layers accumulate. They don't replace each other.

APIs did not remove the need for good interfaces, and conversational products still rely on APIs. AX is the same: an additional layer, built on the business logic you already have, that serves a new kind of user without degrading the experience for the existing ones.

Agent Experience layerAgents acting for people
Natural-language interfacesPeople, via language
Programmatic interfacesDevelopers' code
Graphical interfacesPeople
Shared business logic, data, and rules

See AX in action

One task. Two stores.

An agent is asked to buy a pair of running shoes. Step through the same task against a store designed only for people, and against one that provides an Agent Experience layer.

SimulationScripted to illustrate the difference. No real agent, model, or store is involved, and the timings are illustrative.

Task given to the agent

Buy trail running shoes, EU size 42, under €150, delivered by Friday.

0/7
  1. 01 Load homepage

  2. 02 Search the catalog

  3. 03 Read product details

  4. 04 Select size and add to cart

  5. 05 Proceed to checkout

  6. 06 Retry checkout

  7. 07 Report to the person

Simulated agent reasoning

Run the simulation, or step through it, to see how the same task plays out against a store built only for human eyes.

Event logsimulated
  1. Waiting to start…

Outcome

Complete the run to see the outcome.

Requests
—
Guesses made
—
Errors recovered
—
Outcome verified
—

Who is AX for?

A shared language for the people who build products.

01

Product Designers

Design interactions beyond the graphical interface.

Map user tasks to capabilities, decide where people stay in the loop, and write the language agents will relay to the people they serve.

02

Developers

Build agent-ready capabilities and interfaces.

Turn existing business logic into typed action contracts, publish a manifest, and expose it through your API or an MCP server.

03

Product Teams

Evaluate and improve how agents interact with products.

Assess readiness across six pillars, prioritize the gaps that block real tasks, and track progress with a shared vocabulary.

Open framework

Early, open, and meant to be challenged.

AX Framework is a version 0.1 draft. The pillars, principles, and patterns here are proposals, written to be tested against real products and revised. There is no standards body behind them and no ecosystem yet — that is what contribution is for.

  • Patterns

    Document a recurring problem and a solution you have seen work in production.

  • Implementation examples

    Show how a principle or pattern looks in a real stack, with code.

  • Research

    Share findings from testing products with agents: what failed, what helped, how you measured it.

  • Proposals

    Suggest changes to the pillars, principles, or vocabulary, with the reasoning behind them.