On July 2, Kunlun Tech announced Tiangong 3.2, headlined by a new feature: Skywork Tags. Its core pitch: an AI Agent permanently resides in any group chat on Slack, Feishu, DingTalk, Discord, or Telegram as a team member.
How it works
- Users @Skywork in a chat and bind an SK/API key; the Agent then joins as a regular group member, able to read context and participate in discussion.
- Shared team Agent: one shared instance per channel, with processing progress visible to all members, supporting task handoffs and long-term context accumulation.
- Kunlun's internal A/B experiments claim that after weeks of heavy use by a 100-person team, the shared Agent outperforms privately fine-tuned personal Agents. Kunlun's conclusion: "a team-raised Agent grows far faster than a personally tuned one."
- Comparable product: Anthropic's Claude Tag (June 24), which lets Claude live in Slack channels. Karpathy called this a third paradigm shift in LLM interaction: first the web, then desktop apps, now AI as a persistent entity within organizations.
- A shared Agent's training data is naturally larger — anyone's conversation benefits everyone's future conversations.
- Its context is organization-level — it understands what the team discusses.
- Its feedback is distributed — criticism is a negative sample, approval a positive one.
- Enterprise IT leads: if your team already uses Feishu/DingTalk, Skywork Tags is the lowest-friction Agent onboarding path — no workflow migration needed.
- AI Agent founders: the personal-vs-team Agent route debate is forming; both Claude Tag and Skywork Tags back team sharing.
- Chinese LLM vendors: Kunlun translated model capability into a "Chinese AI employee" — a product template for GLM, ERNIE, Qwen.
- IM platforms: Agents entering chats turns IM from a communication tool into an operating interface, forcing platform strategy questions.
- Data isolation: how does a shared Agent separate personal vs. group context? Privacy boundaries are undefined.
- Pricing: whether it's free for all groups, and tiering for large (>100-person) groups, is undisclosed.
- Multi-Agent collaboration: only one Skywork instance per channel, but enterprises may want finance + HR + engineering Agents working in parallel — a hard product limitation.
- Gap vs. Claude Tag: whether Tiangong 3.2 matches Claude on SWE-bench and Coding Agent benchmarks needs independent evaluation.
- Compliance: in finance, healthcare, and government scenarios, does a third-party Agent in IM chats trigger data export or security reviews?
- https://tech.ifeng.com/c/8uQaxgPPUAl
- https://www.chinaz.com/ainews/29331.shtml
- https://news.aibase.com/zh/news/29331
- https://www.163.com/dy/article/L0R263FO0556I485.html
The underlying logic
Most Agent products ask users to move context to a new platform (Skywork's own platform, Manus desktop, Devin sandbox, etc.). Skywork Tags inverts this: "don't move the context, move the Agent" — the Agent enters the user's existing work environment and inherits all existing collaborative context, including history, files, and member roles.
This is a product-philosophy choice rather than a technical breakthrough, and it splits the field into two paths:
1. Agent as personal tool — Cursor, Claude Code, Devin, Manus: personal, private config, context locked to the user's account. 2. Agent as team asset — Claude Tag, Skywork Tags, Notion AI, Lindy Teams: shared instances, chat-resident, long-term memory, trained on team data.
Kunlun is betting Agent value comes from team knowledge accumulation. Their reasoning: