Actionability
Can an agent execute supported actions?
Actions are explicit contracts with typed inputs, preconditions, and predictable effects, not sequences of clicks to be reverse-engineered.
Most digital products expose actions through interfaces built for hands and eyes. An agent can drive those interfaces, but the result is brittle: a redesigned checkout or a renamed button breaks the automation.
Actionability means that the things a product wants agents to do are available as explicit, documented operations. Each operation states what it needs, what must be true before it runs, what it changes, and whether it is safe to repeat.
This does not require a separate product. Often it means formalizing the operations a UI already performs and making them callable through a contract that is stable across visual redesigns.
What good looks like
Signals
- Each action has a stable name, typed input schema, and documented output.
- Preconditions are declared and checked before execution.
- Side effects are described, including whether the action is reversible.
- Idempotency is supported for actions that create or charge.
- Multi-step tasks are modelled as explicit workflows with named steps.
Anti-patterns
- Actions that only exist as UI event handlers.
- Inputs validated only by client-side form logic.
- Operations whose effects differ based on hidden UI state.
- Retrying a request creates a duplicate order or charge.
Evaluation checks
Use these questions when reviewing a task. They feed the AX evaluation.
- A1Are core tasks available as named actions with typed inputs?
- A2Are preconditions validated server-side with specific error codes?
- A3Do state-changing actions accept an idempotency key?
- A4Are multi-step tasks resumable at each step?