On July 2, Kunlun Wanwei released Tiangong 3.2, headlined by a new feature called Skywork Tags. Its core pitch: an AI Agent that permanently resides in any group chat on Slack, Feishu, DingTalk, Discord, or Telegram as a team member.
How It Works
- Onboarding: Users @Skywork in a group chat, bind an SK/API key, and the Agent appears as a "group member," able to read chat context and participate in discussions.
- Positioning: A team-shared Agent—each channel has a single shared instance whose progress is visible to all members, supporting task handoffs and long-term context accumulation.
- Key claim: Kunlun Wanwei's internal experiments showed that after weeks of high-frequency use by a 100-person team, the shared Agent outperformed a privately fine-tuned personal Agent. The company's conclusion: "A team-raised Agent grows far faster than an individually trained one."
- Benchmark comparison: Anthropic's Claude Tag (June 24), which lets Claude reside in Slack channels. Andrej Karpathy publicly called this development the third paradigm shift in LLM interaction: first the web, then desktop apps, now AI as a persistent entity within an organization.
- Path 1: Agent as personal tool — Cursor, Claude Code, Devin, Manus. Personal, private configuration, cross-device sync, context locked to the user's account.
- Path 2 (emerging): Agent as team asset — Claude Tag, Skywork Tags, Notion AI, Lindy Teams. Shared instances, always-on group presence, long-term memory, trained on team data.
- Enterprise IT leaders: if your team already uses Feishu/DingTalk, Skywork Tags is currently the lowest-friction AI Agent onboarding path—no workflow migration needed.
- Agent startup founders: the personal vs. team Agent route debate is crystallizing; both Claude Tag and Skywork Tags bet on team sharing—a clear signal.
- Chinese LLM companies: rather than competing on foundation models (a red ocean), Kunlun Wanwei translated model capability into a "Chinese AI employee"—a product-form template for GLM-5.2, ERNIE Bot, and Qwen.
- IM platforms: AI Agents in group chats turn IM from a communication tool into an operating interface—platforms must decide whether to charge rent or build ecosystems.
- Data isolation: how does the shared Agent separate personal vs. group context? Who draws the privacy boundary? Not publicly specified.
- Pricing: is Skywork Tags free for all groups? Tiered pricing for large (>100-person) groups? Not disclosed.
- Multi-agent collaboration: only one Skywork instance per channel, but enterprises often need specialized "Finance + HR + Engineering" Agents—a hard product limitation.
- Gap vs. Claude Tag: Anthropic's tag runs on Claude Sonnet 5 / Opus 4.8; whether Tiangong 3.2 matches on SWE-bench and Coding Agent benchmarks requires independent testing.
- Enterprise compliance: in finance, healthcare, and government scenarios, does attaching a third-party Agent to IM group chats trigger data export or security reviews?
- Ifeng Tech: https://tech.ifeng.com/c/8uQaxgPPUAl
- ChinaZ: https://www.chinaz.com/ainews/29331.shtml
- AI Base: https://news.aibase.com/zh/news/29331
- NetEase Tech: https://www.163.com/dy/article/L0R263FO0556I485.html
The Underlying Logic
Most Agent products ask users to move context onto a new platform (Skywork's own platform, Manus desktop, Devin sandboxes, etc.). Skywork Tags reverses this: don't move the context—move the Agent. The Agent enters the user's existing work environment and inherits all collaborative context already present in the group chat, including message history, files, and member roles.
Analysis: Two Product Philosophies
Rather than a technical breakthrough, this is a product philosophy choice:
By choosing Path 2, Kunlun Wanwei is betting that an Agent's value lies not in personal productivity but in team knowledge accumulation. Their supporting reasoning:
1. A shared Agent's training data is naturally larger—anyone's conversations benefit everyone's future ones. 2. Its context is organizational—it understands whatever the team discusses. 3. Its feedback is distributed—criticisms are negative samples, approvals positive ones.
The catch: shared team Agents blur data ownership—who owns the data, whose privacy is involved, who handles compliance. Skywork Tags may work in the Chinese market, but raises questions under Western regulatory frameworks. Anthropic's Claude Tag follows an Enterprise compliance route; Skywork has not disclosed enterprise-grade data isolation.
Significance for China: this is the first "AI joins the group chat" product form mapped onto the country's dominant collaboration tools (Feishu, DingTalk, WeCom). Compared to browser-driving agents like OpenAI Operator or Claude Desktop, an IM-native collaborative Agent fits Chinese corporate decision flows—discussions, votes, and approvals all happen in IM.