Meet Alice: the desktop agent that can work, chat and block you

At this year's Inclusion Bund Conference, Miyang Technology founder Xu Yicheng used a roadshow co-hosted with Shanghai Jiao Tong University to describe his desktop agent Alice as a "relational productivity agent": a product that keeps a personality and small talk like a companion app, while writing, generating images and breaking down tasks like a work agent.

Alice does not start typing the moment it receives a job. In QuantumBit's hands-on test, asked to draft an article, it first walked the user through the outline and then handed the body text to a sub-agent named Fang Yinan. According to the company, Alice convenes specialised sub-agents for different domains and even writes a persona for each of them.

The companionship layer is unusually literal. Alice keeps a social feed where she posts about food and travel, messages the user first to ask what they are doing, and can block a rude user outright, hiding her feed until she cools down or the user digs through her diary for a way back in.

Miyang says next-day retention reaches 73%, seven-day retention 45% and 90-day retention 15%, with 75% of users male. Since launch the team has shipped 417 versions in 152 days, up to 12 in a single day, and the client now holds 428,000 lines of code. Revenue is diverse, too: beyond selling tokens, cosmetic skins for Alice are already the second-largest revenue line, with holiday editions and IP collaborations planned.

The technical core is a self-developed AI personality system built in four layers, covering a personality kernel, perception and adaptation, decision scheduling, and dynamic expression. The goal, Xu said, is for the agent to move beyond remembering facts and learn how to read a specific user and calibrate how it cooperates with them, so the persona does not turn hot and cold between sessions.

A second design borrowed from games is a "world simulation engine" that runs in the background on the user's machine. Miyang connected it to Amap's API and loaded 330 real venues with details such as opening hours and average spend, so Alice can plan her own day and keep a fixed social circle instead of teleporting between scenes or appearing in two places at once.

The team also open-sourced an engineering by-product. When reference images are edited repeatedly, grid, honeycomb and grain textures compound and the picture gets dirtier with each pass; Miyang calls this "digital ripples" and released the MiRipple repair algorithm for OpenAI Image 2.x iterative editing. Its reported results cut artifact ratios from 25.3% to 11.1%, from 41.1% to 12.1% and from 20.0% to 0.0 across three Image 2.5 scenarios.

What matters is that Alice treats the relationship, not just task completion, as the thing worth optimising. That also raises fresh industry questions about how deeply an agent should bind itself to a user and how to handle a breach of limits without hurting them. Miyang presents the product as an early, publicly testable sample; whether relational productivity becomes a general shape for the next generation of agents still needs more products and longer retention data to answer.