agentobjectmodel.org

Agent Object Model (AOM)™ · Spec and Tools

Version License

A task-centric, entity-driven JSON standard that gives AI agents a clean, automation-aware view of any web page or surface — with zero layout noise, clear tasks, and explicit policies.

Get the project: Clone this repo or download ZIP — spec, schemas, examples, and validators in one package. Latest release: v0.1.0.

Agent Object Model and AOM are trademarks; registration has been filed. See Trademark Notice. This repo is the canonical source for the AOM spec and reference tooling (agentobjectmodel.org).

Table of Contents

The Problem

AI agents browsing the web today face critical challenges:

What agents see:

This leads agents to:

Example: A simple login form

What the agent gets (HTML):

<form id="login_form_a8f3" class="needs-validation">
  <input type="email" id="email_input_x2k9" required>
  <input type="password" id="password_x7k1" required>
  <button type="submit">Log In</button>
  <a href="/account/change-password" class="link-secondary">Change password</a>
</form>

Result: Agent wastes tokens, unclear what happens after submit or what is safe to do.

The Solution

Instead of raw HTML, agents receive a clean, semantic AOM JSON document with exactly what they need:

{
  "aom_version": "0.1.0",
  "surface_id": "app:auth:login",
  "surface_kind": "screen",
  "automation_policy": "allowed",
  "purpose": {
    "primary_goal": "Authenticate the user and establish a session.",
    "user_roles": ["guest", "anonymous"]
  },
  "tasks": [{
    "id": "login",
    "label": "Sign in",
    "description": "Submit credentials to authenticate.",
    "default_action_id": "submit_login",
    "input_entities": ["LoginCredentials"]
  }],
  "entities": {
    "LoginCredentials": {
      "schema": {
        "username": { "type": "string", "required": true },
        "password": { "type": "string", "required": true }
      }
    }
  },
  "actions": [{
    "id": "submit_login",
    "label": "Log In",
    "category": "mutation",
    "input_entities": ["LoginCredentials"]
  }]
}

Result: Agent sees purpose, tasks, entities, and allowed actions. No layout noise; automation policy can restrict to AOM-only (READY / allowed) or allow more (OPEN / open).

AOM’s answer:

See spec/v0.1.0/README.md for the full formal model.

Quick Start

Clone or download this repository (see Get the project above), then from the repo root run the commands below. You need either Python 3 or Node 18+ for the CLI.

0. Explore examples

# View an example surface (or open the file in an editor)
cat examples/v0.1.0/login-single/login.aom.json

1. Validate a single input and output

python aom.py validate input --file examples/v0.1.0/login-single/login.aom.json
python aom.py validate output --file examples/v0.1.0/login-single/outputs/_login.success.output.json

2. Run bulk validation across examples

python aom.py validate all --examples-dir examples/v0.1.0
node   aom.mjs validate all --examples-dir examples/v0.1.0

3. Run one demo-agent case

python aom.py demo run --lang python --folder v0.1.0/login-single --test-case _login.success.output
node   aom.mjs demo run --lang node   --folder v0.1.0/login-single --test-case _login.success.output

Sample agent output (single-shot):

{
  "mode": "single",
  "action": {
    "action_id": "submit_login",
    "params": {},
    "priority": 5
  },
  "meta": { "done": true, "confidence": 0.95 },
  "thought": "Proceeding with default task action.",
  "result": { "ok": true, "user_id": "user1234" }
}

For the full matrix of CLI and script-level commands, see COMMANDS.md.

Core Concepts

Concept What it represents Where to learn more
Surface JSON description of a screen: purpose, tasks, entities, actions, state, navigation, signals spec/v0.1.0/README.md (aom-input-schema.json)
Output Agent’s response: thought, chosen action, result, meta spec/v0.1.0/README.md (aom-output-schema.json)
Automation policy Rules for automation: forbidden | allowed | open spec/v0.1.0/README.md, spec/well-known-policy.md
Site policy Well-known JSON for site-wide automation policy /.well-known/aom-policy.json examples
Signals & test cases Built-in feedback and test cases for each surface spec/v0.1.0/README.md (signals.test_cases)

Surface sections

Section What it tells the agent
purpose Why does this screen exist? What is the user’s goal?
tasks What workflows are available? What are the steps?
entities What data is in play? What are the schemas?
actions What can the agent do? What are the inputs/outputs?
state Where are we in the workflow? What is the current context?
navigation Where can the agent go next? What are the transitions?
signals Are there errors, warnings, or confirmations?

Key principles

  1. Task-centric — Organized around user goals, not UI layout.
  2. Entity-driven — Data structures explicit and typed.
  3. Action-oriented — Clear definition of what can be done.
  4. State-aware — Workflow position and context tracked.
  5. Layout-free — No CSS, coordinates, or visual information.
  6. Semantic-only — Pure meaning, zero presentation noise.
  7. Automation guardrails — forbidden / READY (allowed) / OPEN (open) control how agents may use the surface.

Output Modes

AOM supports two execution patterns for agent behavior:

See spec/v0.1.0/README.md and aom-output-schema.json for the full output contract.

Repository Structure

agentobjectmodel.org/
├── spec/
│   ├── v0.1.0/                    # Schemas, templates, well-known policy
│   │   ├── aom-input-schema.json
│   │   ├── aom-output-schema.json
│   │   ├── site-policy-schema.json
│   │   ├── README.md
│   │   └── templates/site-policy/
│   └── well-known-policy.md
├── examples/
│   └── v0.1.0/                   # login-single, ecom-flow, _forbidden-page-template, demo-agents
├── Tools/                        # Python + Node: validate, create-outputs, testing
├── static/                      # Badges, USAGE.md, TRADEMARK-NOTICE.md
├── .well-known/                 # Example site policy JSON
├── aom.py                        # Python CLI
├── aom.mjs                       # Node CLI
└── COMMANDS.md                   # Full command reference

Key entrypoints: aom.py, aom.mjs, and COMMANDS.md.

Getting Started

For agent developers

Goal: Build agents that consume AOM surfaces.

  1. Read the spec and schemas.
  2. Run the examples and demo agents.
  3. Validate surfaces and outputs with aom validate input / aom validate output / aom validate all.
  4. See COMMANDS.md for the full CLI surface.

For site and page publishers

Goal: Make your site AOM-ready.

  1. Serve /.well-known/aom-policy.json (see well-known-policy.md and spec/v0.1.0/templates/site-policy/).
  2. Embed AOM or JSON-LD in pages where you want agent automation.
  3. Reference plugins (WordPress, Next.js, Nuxt, Gatsby, Shopify, static-site) and the browser extension are available at aom.tools; direct downloads: plugins and extension in Community and Support.
  4. Optional: use badges to signal automation policy to users.

In automation_policy: "allowed" (READY) mode you decide which actions the agent can see and perform: if you do not include a password reset or change-password flow in the surface’s tasks / actions, a conforming agent cannot invoke those operations or see the associated sensitive state. You can keep higher‑risk flows on separate surfaces with stricter policy or A2H requirements.

For implementers

Goal: Validators, create-outputs, or other tooling.

  1. Read the schemas (aom-input-schema.json, aom-output-schema.json, site-policy-schema.json).
  2. Run the Tools (validate, create-outputs) via CLI or scripts.
  3. Use COMMANDS.md for all commands and parameters.

Examples

Example Purpose Location
Login single Single-shot sign-in surface (READY guardrails) examples/v0.1.0/login-single/
Ecom flow Multi-step checkout flow examples/v0.1.0/ecom-flow/
Forbidden page template Page-level no-automation examples/v0.1.0/_forbidden-page-template/

Demo agents (Python + Node) that consume surfaces and produce conformant outputs live in examples/v0.1.0/demo-agents/.

See examples/README.md for details and validation instructions.

Tools and Validators

Quick start: run python aom.py --help or node aom.mjs --help from the repo root. Full reference: COMMANDS.md and Tools/README.md.

Roadmap

See CHANGELOG.md for release history.

Contributing

This repo is the reference implementation for the AOM spec. Contributions are welcome, especially:

How to contribute:

See CONTRIBUTING.md and spec/v0.1.0/README.md for versioning and compatibility guidance.

License

MIT License — see LICENSE. You may use, modify, and redistribute in commercial and non-commercial projects.

Why AOM?

Community and Support

This site is open source. Improve this page.