A tutorial on Juejin this week showed: send "help me package this" to Aco AI's Agent (an AI that autonomously operates a computer), and it directly operates Jenkins to complete version packaging—and the real key is distilling that conversation into a reusable "AI colleague."
We noticed: the Agent concept Silicon Valley has been pitching for half a year has, for the first time, landed in a form Chinese developers can understand—on "packaging," the most common repetitive grunt work in enterprises.
What this is
The whole flow has four steps: connect a local computer → use natural language to let AI complete one packaging run → distill the successful conversation into a task-specific "AI teammate" → next time @ it for direct invocation. Aco AI belongs to Aco, one of China's earlier desktop Agent companies.
What sets it apart from regular chat AI is step three—it packages the rules learned from the conversation into a reusable Agent. This is exactly the direction Anthropic's Computer Use, Zhipu's AutoGLM, Manus, and others are betting on: moving AI from "can talk" to "can work, and can remember how to work."
Industry view
The optimistic voices come from front-line developers. Many in the tutorial's comment section sigh, "finally don't have to re-explain the Jenkins environment every time." A test engineer wrote: "Last year we were still debating whether Agents could land. This year I'm already using one to run daily packaging."
But sober warnings are worth noting. First, these "AI teammates" are highly dependent on vendor ecosystems—Aco-generated assistants can only run within the Aco system; migration cost and vendor lock-in are real issues. Second, they offer limited help for complex tasks that require judgment; what can be codified is precisely "relatively fixed workflows," and much of enterprise repetitive labor is not that. Third, permission and audit gaps: AI directly executes local commands—if something goes wrong, who's responsible? Engineers and compliance departments don't have answers yet.
Impact on regular people
For enterprise IT: we recommend piloting first with one "done weekly, fixed steps, mistakes not fatal" process—such as scheduled packaging or data export—get it working before deciding whether to scale. Don't bet on company-wide automation from day one.
For individual careers: in the next 1-2 years, "can you distill repetitive work into reusable AI workflows" will become a new bonus skill—like when Excel macros first emerged. Not required, but those who use it pull noticeably ahead in efficiency.
For consumer markets: today this mainly targets developers and ops, far from ordinary consumers. But if Xiaomi, Huawei, or similar embed comparable capabilities into their system assistants, by next year we may see "phone automatically orders coffee, checks logistics" actually work.