Skip to content
AX Framework
Principles

Ten principles for agent experience

Principles describe the qualities a product should have, independent of technology. Each one includes a good and a bad example, implementation guidance, and criteria you can evaluate against.

AX proposalProposed by AX Framework. Not an externally ratified standard.
  1. 01

    Explicit Capabilities

    DiscoverabilityActionability

    A product should declare what an agent can do with it, in a machine-readable form, rather than leaving the agent to infer capabilities from the interface.

  2. 02

    Structured Context

    Comprehension

    Information an agent needs to make a decision should be available as typed, structured data with explicit units, relationships, and constraints.

  3. 03

    Predictable Actions

    Actionability

    The same action with the same input in the same state should produce the same effect, and that effect should be documented before the action is taken.

  4. 04

    Transparent State

    ComprehensionFeedback & Observability

    An agent should be able to read the current state of anything it is working with, directly, rather than reconstructing it from a sequence of past actions.

  5. 05

    Clear Permissions

    Trust & Control

    What an agent is allowed to do should be explicit, scoped to the task, visible to the person who delegated it, and revocable at any time.

  6. 06

    Verifiable Outcomes

    Feedback & Observability

    Every action should return evidence of what it did, and that evidence should be independently checkable later.

  7. 07

    Graceful Failure

    Recovery & ResilienceFeedback & Observability

    When something goes wrong, the product should explain what happened, whether it is safe to retry, and what the agent can do next.

  8. 08

    Minimal Ambiguity

    ComprehensionRecovery & Resilience

    Each term, action, and value should have one meaning. When a request could mean more than one thing, the product should ask rather than guess.

  9. 09

    Human Oversight

    Trust & Control

    People should be able to see what agents are doing on their behalf, approve consequential actions, and intervene at any point.

  10. 10

    Progressive Autonomy

    Trust & Control

    The degree of independence granted to an agent should increase with demonstrated reliability and decrease with the stakes of the action.