If an application gains an agent, its interface acquires a new responsibility: showing which actions the system is allowed to take and how a person can inspect the result. A chat box is a way to request work. It is not, by itself, a complete model of permission or accountability.

Why it matters: A fluent explanation of an action can be mistaken for evidence that the action was authorized or completed correctly.

Separate discovery from authority

Apple’s App Intents framework describes how application actions and data can be exposed to supported system experiences. Anthropic’s agent guidance describes a different layer: systems that can choose steps and tools while working toward a task.

These concepts can be discussed together without treating them as one product or assuming a particular integration. Making an action discoverable does not answer every question about when an agent should be permitted to invoke it.

A team needs an explicit boundary around the objects, accounts and operations in scope. The phrasing of a user’s request should not silently expand that boundary.

Make confirmation carry useful information

For a hypothetical agent preparing an external message, confirmation should show the recipient, the content and the intended send action. A generic “continue” button provides less information than the decision requires.

Some operations can be performed safely within a narrow standing permission; others deserve review at the point of execution. The choice depends on consequences and reversibility, not merely on how confident the model sounds.

When an AI Product Needs a Human in the Loop examines what makes a review meaningful. The important feature is the reviewer’s ability to understand and change the proposed outcome.

Verify the state after the story

Anthropic’s evaluation guide distinguishes the agent’s transcript from the outcome in its environment. A claim that a file was updated is different from a check that the correct file contains the intended change.

Products should retain an intelligible action history and a way to resolve partial completion. If one step succeeds and another fails, the final message should describe that state accurately.

The headline asks a hypothetical question. It does not claim every app currently has an agent. As these capabilities are added, the valuable design work will often be unglamorous: bounded access, useful confirmations and an honest account of what actually happened.

Reporting & Sources

Architecture and design analysis, reviewed September 26, 2026. The message-sending scenario is hypothetical.