Agent Accessibility Guidelines (AGAG)

Any meaning conveyed visually (color, icon, position, shape, animation) must also be exposed as text or semantic data.

Agents and assistive tech cannot reliably infer meaning from visuals alone.

Interactive elements must expose consistent, machine-detectable roles, names, and purposes.

Charts, diagrams, maps, and instructional visuals must include structured descriptions or data equivalents.

Every input, select, textarea, toggle, and custom form control must have a programmatically associated label.

Agents and assistive tech need unambiguous control identity.

Related fields must be grouped using semantic containers such as fieldset and legend, or equivalent accessible grouping structures.

Grouping exposes task logic and prevents cognitive fragmentation.

Tables must use proper header cells and associations so agents and assistive tech can interpret row and column meaning.

Tables must not be used for layout, and visual layout structures must not impersonate tables unless they expose true tabular semantics.

Bots and assistive tech need to know whether content is spatial decoration or relational data.

Audio and video must expose whether captions, transcripts, and audio descriptions are available.

Pages must expose core landmarks such as header, navigation, main content, search, complementary, and footer where applicable.

Pages must declare their primary language, and major language switches must be marked where feasible.

Dynamic updates must be announced via machine-readable mechanisms.

Errors must include clear, actionable next steps.

The system must expose where the user or agent is within a process.

Every validation error must be programmatically associated with the field that caused it.

Required status must not rely only on asterisks, color, or layout.

Play, pause, mute, captions, speed, and transcript access must be discrete controls, not gesture-only or hover-only.

Dialogs, alerts, and popovers must announce why they appeared and what the next choices are.

Lazy-loaded, streaming, or asynchronously refreshed content must expose loading state, completion state, and failure state.

If a session is about to expire, the system must announce this in time for action and expose a path to extend or save progress.

Any gesture-based action must have a discrete alternative.

Focus navigation must follow the logical order of the task.

No required action may depend on timing, holding, or rapid input.

Users and agents must be able to skip repeated navigation and jump to primary task content.

Links and buttons must expose destination or action meaning without requiring visual inference.

When a modal or dialog opens, focus must move into it, remain meaningfully constrained, and return to the invoking control when closed.

Authentication and abuse prevention must not require image interpretation, press-and-hold behavior, cursor path tracing, or other mechanisms that exclude legitimate assistive agents or disabled humans when no accessible alternative exists.

Avoid unnecessary DOM complexity that obscures meaning.

Expose structured data or APIs for repeatable tasks.

Buttons and actions must clearly represent their real outcome.

Time-sensitive information must expose source and date.

Instructions required for task completion must be visible, textual, and machine-detectable. They must not depend on inference from layout, color, or prior hidden knowledge.

Medical, financial, or irreversible actions must require confirmation.

The system must expose what an agent is allowed to do.

Users must be able to interrupt or override agent actions.

A system must distinguish between harmful automation and authorized delegated assistance, and must expose a compliant path for trusted assistive or authorized agents.

The web cannot treat all bots as abuse if some bots are literally performing accessibility support.

For multi-step tasks, the system must preserve or restore in-progress state wherever possible, especially for high-stakes processes.

Loss of state is both an accessibility failure and an agentic failure.