openJiuwen's Hyper-Agent lets AI workers live inside Feishu and DingTalk chats

What openJiuwen released

openJiuwen is an open-source agent community backed by Huawei. It has published Hyper-Agent, a design for running many AI agents together as a "swarm." As pandaily.com reported, based on a community release first carried by IT Home, the aim is to move agents out of the private chat window and into the shared channels where teams already work.

The features ship in a product called openJiuwen WorkSwarm. Its code is public on GitHub under the name JiuwenSwarm, so developers can inspect it and build on it.

An agent that outlasts the conversation

In a typical chatbot, the conversation is the main thing and the AI is whatever answers inside it. Hyper-Agent flips that. The agent is the lasting object, with its own identity and memory. A session, meaning one thread of conversation, becomes a resource the system hands out, runs side by side and picks back up later with context intact.

Two pieces of plumbing make this work. A gateway connects messaging apps and a web interface to the system. An AgentServer decides which agent handles which piece of work. The community says this extends its earlier work on sessions that can be resumed, and it lets one agent juggle several jobs across several outside group chats.

Long jobs hit a familiar limit: a model can only hold so much text in view at once, which is called its context window. When an agent nears that limit, it opens a fresh window and brings a short note covering goals, progress and next steps. If it needs a detail, it looks in the archived history.

Hiring an AI colleague

WorkSwarm includes an employee center where a user sets up an agent, either from an expert template or from a blank page. The profile covers the agent's duties, working style and what a finished delivery should look like. It also covers skills, connectors to other tools, knowledge sources and a dedicated email inbox.

The agent also has rules for when to speak. In a passive mode it answers only when someone mentions it. In an adaptive mode it judges from its role and the flow of conversation whether to stay quiet, reply or jump in. An activity panel shows what each agent has been doing, how many tokens (the units models are billed in) it has used, and its recent response times.

Crossing chats, with permission

One agent can work across office apps such as Feishu and DingTalk, both widely used workplace messaging tools. It might receive a request in one and send the finished result to the other.

Moving information between groups needs a human sign-off. The user reviews the source group, the recipient and the content, then grants permission once or permanently. That permission can be revoked later. This is the sort of control a team would want before letting software carry information between rooms.

• Check where the information comes from.
• Check who will receive it.
• Check exactly what is being shared.

Teams, laptops and outside experts

Team members can run WorkSwarm on their own computers. Agents on a cloud server and agents on personal PCs can then cooperate in one group, with the local ones using files and tools the owner has authorized.

For bigger projects, a group chat called SwarmEngine puts people, persistent agents, single experts and expert teams in the same space. An expert team has a leader that accepts tasks, reports progress and organizes the final delivery while the other members work in the background. Experts can also be third-party agents such as Codex and Claude Code, mixed with the in-house digital employees.

The community lists several things still to come: agents coordinating personal and cluster computing resources, active learning, and recovery from failures during longer collaborations.