On July 27, Alibaba's new Agent product "Qwen Work" (QwenWork) started a closed desktop beta — the first time Alibaba has fused three of its agents — QoderWork (coding), Wukong (general), and MuleRun — into a single client using an ACS collaborative architecture. Under the hood is the latest flagship Tongyi Qwen model, which by product positioning sidesteps lightweight helpers like DingTalk AI and Feishu Miaojí, and goes head-to-head with Claude Code and Manus at the "deliver complete artifacts" desktop Agent tier. The first 12 vertical industry kits of Qwen Work (HR / finance & tax / legal / investment research / e-commerce, etc.) cut office scenarios into reusable workflows, and combined with the independent client + full DingTalk IM integration + interactive web generation, this may be the first time Alibaba has run the full "LLM → Agent → vertical knowledge" pipeline end to end. But the desktop-client-first cadence also exposes unfinished business: the web client isn't wired up, Agent orchestration stability is unknown, and the pricing tiers (free / ¥98 / ¥198 per month) span very different capability levels — so the test is whether the closed beta can pull real enterprise pass-rates up to something usable.