Skywork Tags, the new feature in Kunlun Wanwei's Tiangong 3.2 AI assistant, runs directly inside Slack, Feishu, DingTalk, Discord, and Telegram.
The next architectural fight in enterprise AI is not about which model is smarter. It is about who controls the persistent context layer. Kunlun Wanwei (昆仑万维), the Beijing-based AI company behind the Tiangong/Skywork model family, just put a stake in the ground on the same side as Anthropic: AI agents should accumulate memory inside the chat tools teams already use, not in a separate workspace users must visit.
Kunlun Wanwei released Tiangong 3.2 with a feature called Skywork Tags. The feature adds a bot-style presence for the AI inside the chat platforms where work already happens: Slack, Lark/Feishu, DingTalk, Discord, and Telegram. Kunlun Wanwei's release explicitly names Anthropic's Claude Tag for Slack as the direct reference point and positions Skywork Tags as a comparable but expanded cross-platform alternative. The two vendors are not formally partnered. They are aimed at the same architectural idea.
That idea — anchoring AI memory inside the collaboration layer rather than a separate agent workspace — has attracted the entire enterprise AI industry. VentureBeat, in its independent coverage of Claude Tag's launch, described the Slack channel as "the most fiercely contested real estate in enterprise AI" and mapped the competitive landscape: Salesforce has transformed Slackbot into a full-spectrum enterprise agent with more than 30 new capabilities; OpenAI introduced Workspace Agents that operate inside Slack alongside Google Drive, Microsoft apps, and Salesforce; Perplexity deployed its enterprise agent with direct Slack integration; Cognition's Devin has used Slack as a primary interface since its early days. Independent analysis from HyperFRAME Research described the pattern as "an architectural migration away from isolated chat interfaces toward persistent, multiplayer agent fabrics," arguing that Anthropic is making "a land grab to dominate the contextual control plane of enterprise collaboration."
Kunlun Wanwei's positioning is explicitly inverse to the prevailing pattern. Most "AI agent" launches of the past year asked teams to bring their documents, code, tickets, and conversations into a new agent-native workspace. Skywork Tags inverts the direction: the agent moves into the team's existing channel, not the other way around. Chinese tech press carried the same framing across multiple outlets, with ifeng and itbear echoing the same release within hours.
The company identifies three concrete benefits of that placement. Transparency: the agent's progress is visible to everyone in the channel rather than hidden in a private tab. Async handoff: a teammate can pick up the work where the agent or another user stopped, with the full trail still in the chat. Compounded context: the agent accumulates long-running memory of the channel's work instead of resetting each session. Translated into vendor language, the AI is positioned as a colleague who joins the channel rather than a tool that demands a new tab.
That vendor metaphor is borrowed framing. Kunlun Wanwei's release invokes a line attributed to AI researcher Andrej Karpathy describing AI as a persistent organizational entity, the "third interaction paradigm" after web and desktop apps. The framing is Karpathy's, not the vendor's, and the release paraphrases rather than quotes him on the record. It is useful scaffolding for an architectural argument, not an established fact.
The hidden mechanism underneath is a reversal in who owns the switching cost. The agent industry's portability pitch has been that personal AI assistants would "follow the user" across apps and devices. If agents instead accumulate persistent memory inside Slack, Feishu, and DingTalk, the lock-in runs the opposite direction: the chat platform operator becomes the persistent memory custodian, and the agent vendor's portability story quietly evaporates. HyperFRAME Research put the same point in analytical terms: the chat platform becomes the "contextual control plane," and Anthropic — like Kunlun Wanwei — is betting that inserting its agent into that layer is worth more than owning a separate agent workspace. The same release copy that frames the move as liberating teams from migration also, in practice, hands the chat platform the long-term context layer.
Kunlun Wanwei offers one piece of internal evidence for a shared design over a personal one. According to the company's release, when its own teams ran a shared team bot in parallel with an individually tuned personal bot, the shared version overtook the personal one in usage within two to three weeks. The reader should treat this as a vendor's anecdotal self-report from a single company's internal experiment, not a published benchmark, and not a generalization beyond what the source actually claims.
The architectural question neither vendor's release answers is whether "shared" is actually "better." The Kunlun Wanwei release positions shared agents as more transparent and more useful for handoff, but it does not address three concrete risks. Context fragments across Slack, Feishu, DingTalk, Discord, and Telegram: a team running on three of these platforms ends up with three separate agent memories and three audit trails. Governance and audit gaps open up when the agent acts inside a third-party chat, since compliance teams that today review model activity inside a vendor's own workspace have to negotiate new log access with the chat operators. And the conflation of "shared with the channel" with "useful to the channel" is not addressed: an AI that sees every message is also an AI that knows every message, and the vendor's release does not address the privacy implications of an agent reading the whole channel by default.
The convergence Kunlun Wanwei claims to be happening, with Anthropic's Claude Tag for Slack and its own Skywork Tags arriving on the same architectural direction in the same window, suggests the bet has cross-vendor support. The interesting fight now is not which model wins, but which chat platform becomes the persistent context layer underneath. Slack, Feishu, and DingTalk are the immediate candidates named in the release itself. The vendors' framing sells the move as AI coming to where teams already work. The mechanism underneath is that the chat platform quietly becomes the new lock-in surface for AI memory.