What happens when a voice assistant keeps working after you walk away?

A new system prototype called Richard separates the conversation you have with your phone from the work the phone does afterward. You can leave the chat, come back later, and the task is still there — revised, finished, or waiting for you.

The prototype has a name and a purpose

Richard is a research prototype built by Xinyang Chen and submitted to arXiv on 30 September 2026. The paper sits under Human-Computer Interaction and Artificial Intelligence. Its goal is to let mobile terminals provide application and network services while supporting users' control over their attention. Voice-first interaction is organized around requests and delegated tasks. Users can leave a conversation and later inspect, revise, and retrieve the work. The system manages voice sessions, task execution, and result delivery separately. Persistent request records link them together.

Conversation and execution live in different layers

The prototype splits three concerns that most voice assistants mash into one flow. A voice session handles the spoken dialogue. A task execution layer runs the actual work — dedicated service operations and agent tasks. A result delivery layer brings outcomes back when the user returns. Conversation and task views give visual feedback. Request revisions, execution states, and notifications stay attached to the relevant task record. The backend coordinates immediate responses, dedicated service operations, and agent tasks without requiring the user to stay connected. If the client disconnects, the backend continues. Progress and failure feedback appear when the user reopens the task view.

The request becomes a durable object

You can revise it after seeing early results. You can close the app, lose signal, or switch tasks. The backend keeps running.

Tests ran on Android and in a deployed service

The authors examined the design through Android functional records, controlled lifecycle verification, and execution records of a real programming request. Controlled verification reproduced revision, execution after confirmation, and result retention. Deployed-service records showed backend progress and failure feedback after client disconnection. The paper does not say how many devices were used. It does not report user study numbers. It does not name the programming task.

The phone becomes a place you can leave and return to

Current voice assistants mostly demand continuous attention. You speak, they answer, the session ends. Richard changes that contract. The paper does not claim a product launch. It does not compare against commercial assistants. It presents a prototype and the evidence that its core loop works — revision, execution after confirmation, result retention, and feedback after disconnection. For anyone who has ever lost a long voice command because the screen locked, the difference is concrete: the request survives the interruption.