On September 7, Alibaba's Qianwen Office (千问办公) launched a multi-person workspace. You describe who is involved, what work they do and which information the team has to keep. It generates a collaboration app with role-based permissions, a cloud database and an administrative backend. geekpark reports the workspace supports up to 100 people online at once.
That is a different product from "chat produces a document". The output is a place where work gets assigned, records persist and someone administers access.
What the early users did with it
geekpark describes OneX AI community founder Huo Ge building a tourism collaboration platform on it: sales staff, operations coordinators and finance all working on the same orders. The first version took about half an hour to generate, then one night of iteration.
ifanr ran a similar test with TeleAgent, a competing office agent, for a startup called Shixu. It produced a positioning document, a recruitment plan, a two-week action plan, a 6-page PPT and a web prototype. It also changed the beta-user quota from 50 to 30 when asked, and where the academic administration had not confirmed a date, it kept both October 29 and October 30 open rather than picking one.
Why the prompt is the easy part
Every one of those examples worked because the rules lived outside the prompt: in the roles, the retained records, the permissions the generated app carries forward. The speed is real, but generated work stays reliable only when the context is bounded.
woshipm's author, who ran a real project with a small set of agents, describes what that bounding looks like in practice: a local project directory with approved source files and outputs, one accepted meaning for each data field, a decision log that says what was accepted and who owns the next call. Even then it took about 2-4 rounds of debugging before the agent's role-to-capability mapping was usable.
So the checklist before handing a substantial task to one of these workspaces is not about prompt wording. It is: approved sources, named owners, data definitions, an explicit rule for what the agent does when a fact is uncertain, and human confirmation before anything consequential.
The full piece walks through the three Chinese approaches (an HR SaaS CLI project, Qianwen Office, TeleAgent) side by side, with the source articles linked: https://eastofsilicon.com/posts/qianwen-office-turns-prompts-into-governed-workspaces
East of Silicon covers tools and workflows from the Chinese internet for English readers. I run it.
Top comments (1)
"The prompt is the easy part" is the most honest framing I've seen for this category of tool. The hard part is exactly what you describe — approved sources, named owners, data definitions, and an explicit rule for uncertainty before anything consequential runs.
The Huo Ge tourism platform example (30 minutes to generate, one night of iteration) tracks with what we see building automated systems at Black Label. The generation is fast; getting the role-to-capability mapping right is where the real work lives. Two to four rounds of debugging before it's usable is actually optimistic for anything involving cross-role permissions and shared records.
What's interesting about Qianwen Office vs the Western equivalents is the emphasis on organizational structure first — roles, permissions, data retention — rather than starting from the prompt and adding structure later. That ordering matters. Looking forward to the side-by-side piece on eastofsilicon.com.