Shanghai Startup Pitches 'Relational Productivity' With Desktop Agent Alice

Founder Xu Yicheng, who also publishes under the name Luo Xiaoshan, spent the past decade at Seasun, Tencent's TiMi Studios and Kingsoft Office. About half of the AI-native team comes from Shanghai Jiao Tong University and nearly all of its members were born after 2000, according to QbitAI.

The company says Alice's next-day retention rate reaches 73%, with 45% at seven days and 15% at 90 days. Seventy-five percent of its users are men.

In a demonstration reported by QbitAI, Alice did not begin drafting an article immediately. It first walked the user through an outline, then handed the writing to a sub-agent named Fang Yinan, described in-character as a senior editor based in Chengdu. The company says Alice assembles specialist sub-agents for different tasks and writes each one a persona. Users can switch the underlying model, including a recent DeepSeek V4.1 Flash option. Generated images are saved directly into a workspace, and a requested National Day travel plan included choices between Wuzhizhou Island and West Island as well as a sunset stop at Coconut Dream Corridor, with restaurants the user could look up.

Alice also maintains a Moments-style feed of food, travel and city walks, sends direct messages asking what the user is doing, and can cut off access to its feed after a rude remark. QbitAI reported that a block lasted a 20-minute countdown before the agent relented. Each exchange moves an affinity score up or down, and the agent adjusts how it interacts based on how familiar it has become with the user.

Xu described a four-layer personality system: a personality core that keeps the agent's traits stable; a perception layer that reads context, mood and tone; a decision layer that combines long-term memory, task state and personality rules; and a dynamic expression layer that adjusts phrasing and style. Memory records what the agent remembers, Xu said, while the personality system decides how the agent should interpret the user and at what distance to work with them. The report contrasted this with GPT-4o, which it said leaned toward accommodation that raised short-term stickiness without matching long-term user satisfaction.

The company also built a world simulation engine, drawing on Xu's game-industry background. Alice connects to Amap's API and a database of 330 real venues with opening hours and average spending, plans its own daily schedule and keeps a fixed social circle, so users returning to the app can see what the agent did while they were away.

On the image side, the team open-sourced MiRipple, an algorithm aimed at the grid, honeycomb and grain textures that compound when reference images are fed back into iterative AI editing under OpenAI's Image 2.x, an effect it calls digital ripples. The three-step pipeline diagnoses frequency-domain and spatial artifacts, applies targeted filtering, structure-aware smoothing or regeneration, then checks for over-correction. QbitAI reported artifact shares falling from 25.3% to 11.1% in a moss canyon case, from 41.1% to 12.1% in a wisteria tunnel case and from 20.0% to 0.0 in an ice cave case; the paper is posted to arXiv.

Since launch, Mianyang says it has shipped 417 versions in 152 days, with as many as 12 in a single day. The client contains 428,000 lines of code across 1,277 source files, with 983 features implemented.

Revenue comes from token sales and from cosmetic items. Xu said appearance sales are already the company's second-largest revenue category, with holiday limited editions and IP collaborations planned. Free AI education content published under Xiaoshan Academy has passed 20 million cumulative visits, and the product topped Guancha's May app ranking, according to the report.

Mianyang's pitch rests on a squeeze facing the agent market: model capabilities are converging, switching costs for productivity agents are low, and companion apps offer emotional value without recurring deliverables. The report cited an Artificial Analysis scatter plot in which the capability-to-price curve rises steeply below $0.4 per unit before flattening, leaving a $5 gap between GPT-6 and Fable 5.1 at similar capability. Relational productivity agents are meant to combine recurring work with continuity, so that accumulated familiarity raises the cost of leaving.

Open questions remain, including how tightly a user and an agent should be bound and how a product should handle a breach of boundaries without harming the user emotionally. QbitAI presented Alice as an early sample rather than a settled answer. The app is available at alice.fans.