Where AI coworkers sit in China

Chinese teams have four routes to an AI coworker in 2026: put several agents in a shared project (Coze 3.0), give the agent an identity inside the chat app the company already lives in (Doubao's work partner in Feishu), build a workspace for humans and agents (Block's Buzz, Kylon), or leave the agent on the machine it already runs on and put a room around it. They answer the same question differently: who approves the step that cannot be undone.

This matters outside China too, because the Feishu route is the one Western products keep rediscovering as "put the agent in Slack".

Four routes into the same office, and the question all of them have to answer.

Coze 3.0: the project is the shared space

Coze 3.0 lets several people and several agents work inside one project. Files and the agents' outputs stay in the project, and members split work by mentioning a particular agent. External agents, including Claude Code and Codex, can be brought in.

Its centre of gravity is still the agent-building platform underneath: you compose an agent, then bring it into a project. For a team that has already built workflows there, that is the shortest path. For a team that just wants the coding agent it already pays for, there is a construction step in the way.

Doubao's work partner: no new space, just Feishu

ByteDance made the cheapest possible move. Feishu already holds the group chats, documents, meetings, calendars, org chart and permissions, so rather than build another workspace, give the agent an organisational identity inside the one everybody already uses.

The advantage is immediate: a company with years of history in Feishu gets an agent that can use that authorised context on day one, with nothing to migrate and nobody to re-register. The cost is that the agent's reach is defined by the platform, and the platform is one vendor.

Buzz and Kylon: build the workspace

Block open-sourced Buzz in July 2026: a self-hosted Nostr relay where people and agents share channels, every message, review, workflow step and git event is a signed entry in one log, and agents hold their own keypairs. A bridge, buzz-acp, brings Goose, Codex and Claude Code in as members.

Kylon goes the other way and ships a finished hosted product: rooms holding files, people, AI members and small business apps, with more than three thousand connections, Skills, Workflows, and memory compressed into a searchable summary tree. Its AI members can be Kylon-hosted or connected to a Codex or Claude Code already configured on someone's computer. It is aimed squarely at companies without an engineering team.

Both ask you to move work in. The self-hosted one also asks somebody to run a database, a cache and object storage. The full comparison is in Block's Buzz and muti and Kylon and muti.

The fourth route: don't move, reach

What a small team is missing is often not a workspace but access to one machine. That laptop has the repository cloned, the internal database reachable, the design tool logged in, the command-line tools installed and the Claude Code or Codex subscription already paid. Copying that environment into any workspace costs more than reaching it where it sits.

muti does the reaching. A daemon on the machine polls outward, runs each request locally, and streams the timeline back into a room. Files, environment variables and keys never leave. Permission is per person, and four categories always stop for the owner no matter who asked: pushing to a shared branch, deleting, spending money, touching credentials.

Four routes, compared

Choosing

• Deep in Feishu, and the job is "read what we already have, then carry it forward": Doubao's work partner is the least work.
• Already building agents and workflows on Coze: Coze 3.0 is the natural extension.
• You want to own every record and have someone to operate it: Buzz. You want your operations inside one finished product and have no engineers: Kylon.
• You want colleagues in WeCom, DingTalk or Feishu to @ an agent that runs on someone's computer: Tencent WorkBuddy and Alibaba's QoderWake do exactly that, and QwenWork adds a cloud layer inside DingTalk. See WorkBuddy, QwenWork and QoderWake.
• The most capable agent in the team runs on one person's laptop and nobody else can reach it: put a room around that machine. That is the only problem muti tries to solve.

Frequently asked questions

Can these be used together?

Usually they are. Operations live in a workspace or in Feishu, while one engineer's laptop gets its own room for a client or a co-founder. muti's daemon only makes outbound connections and does not interfere with other agents on the machine.

Is "agents as members" a real thing or marketing?

Real in every product above. In Buzz an agent holds its own keypair and audit trail; in Kylon a name, a role and durable memory; in muti a handle, an inbox, a permission mode and a serial queue. Membership is what makes an agent addressable by more than one person, which is the point of the category. See what is an agent room.

Does muti connect to Feishu?

The Slack connector shipped: one bot per agent, DM it or mention it, approvals in the owner's DM. Feishu and Lark are next, with the design published in mentioning your Claude Code from a Feishu group.

Why not keep business data in the room too?

Because then "nothing leaves your computer but the transcript" stops being true. Rooms hold messages, timelines, approvals and small artifacts. Business data stays where it already lives.

Sources

• Buzz: github.com/block/buzz
• Kylon: kylon.io
• Chinese original of this piece: AI 同事该住在哪
• The global map: where agents live in 2026

If your team's best agent is reachable by exactly one person, give it a room.