WorkSwarm: When AI Agents Become a Collaborative Team for Computer Vision

From Solo Assistant to a Working Team

For a while now, the race in office software has been about making AI assistants smarter at single tasks: draft a doc, fill a spreadsheet, schedule a meeting. But real work is rarely a one-shot generation. You gather materials, write, revise, hand off files, sync progress. The messier the job, the more moving parts. When AI starts doing that kind of work, a new question appears: can multiple agents and humans work together like a team—dividing tasks, picking up where someone left off, and actually delivering?

openJiuwen, a collaborative effort from Huawei's 2012 Lab, Huawei Cloud, and device teams, has been exploring exactly that. Their answer is WorkSwarm, a swarm of office agents that brings multi-agent collaboration into personal work. It just landed on the HarmonyOS PC app store, with support for Windows and Mac. The demos are playful but pointed: one has seven agents write, compose, arrange, and perform a song together. Another has two AI writers and a human share the same Word doc, taking turns adding lines to a classic piece of Chinese literature.

Seven Agents Walk into a Studio

The music demo starts with a simple request: a song about a coastal typhoon, with a cinematic, narrative feel. WorkSwarm spins up a team of seven—a team lead, lyricist, composer, arranger, lead vocalist, accompanist, and interlude specialist. Each has a clear role. The lyricist writes words and singing cues. The composer sets melody and structure. The arranger handles instrumentation. The lead vocalist focuses on diction, breath, and emotion. The interlude specialist checks transitions between sections. The accompanist suggests harmonies and dynamics.

They settle on a style—Epic Cinematic Dark Pop—and map out Intro, Verse, Chorus, Bridge, and so on. A first demo appears quickly. Then comes the interesting part: after listening, the team starts critiquing. The lead vocalist thinks the emotion could push harder in spots. The interlude specialist goes section by section. The accompanist weighs in on how finished things feel. They debate the Bridge, the Chorus build, the length of the interlude, where the burst should hit. Those notes feed back into the creation loop.

The version numbers tick up from music-2.0 to music-2.6, and every change is logged. The final deliverable includes twelve files: MP3, lyrics, composition plan, arrangement summary, vocal instructions, accompaniment suggestions, interlude review, version history, delivery summary, and a style prompt. If you want to keep tweaking later or reuse the style for a new track, all the material is right there.

Human on the Swarm, Human in the Swarm

The second demo is quieter but just as telling. Three members: two AI writers and one human. The task: take turns writing new lines for an ancient Chinese essay, one sentence per turn, signing each line with a name. They share a single Word document on the desktop. When it's an AI's turn, it reads the current content, finds where to continue, writes a line, and tags it with its handle. Then it saves, closes Word, reopens it, and pings the next member in the group chat.

The human joins from a phone through an invite. When it's their turn, the phone shows the latest text and the task. They write a line, sign it, and hand the baton back. The Word doc updates on the desktop, and the chat logs the handoff. Those small actions—closing and reopening Word—sound trivial, but they mirror how real work happens. The agents have to understand the writing brief, handle file operations, and communicate progress.

What stands out is that the human isn't a supervisor hovering above the swarm. They're inside it, with a role, tasks, and a voice. They can step in to write, or step back and let the AIs take turns. It's a natural way to work alongside agents, not just command them.

Four Things That Make It Work

Under the hood, WorkSwarm relies on a few key capabilities. First, it can assemble a team on its own. Given a goal, it matches roles, breaks down the workflow, and organizes collaboration. You don't have to summon each tool or manually shuttle outputs between them. Second, it shares context across the team. Lyrics, arrangement notes, demo versions, Word documents—all become shared assets. Each agent builds on what came before, cutting down on repeated explanations and lost information.

Third, it moves beyond chat. Agents don't just generate text; they read and write files, operate applications, and deliver finished work. The outputs live in real tools and real documents, not in a conversation window. Fourth, the process is transparent. You can see who did what, what's pending, what reviews came back, and how versions evolved. And successful workflows can be saved as a Swarm Skill, so a team that figures out a good way to work once can reuse it next time.

From Batch Reports to Code on the Fly

The demos are charming, but the everyday use cases are where this gets practical. Picture getting an urgent briefing task on a train. You open a messaging app on your phone, fire off a few words, and a team on your PC starts pulling together the materials. You're still traveling, but the work is already underway. If the outline misses the mark, you send it back. If a technical detail needs a closer look, you tag an expert agent. Halfway through, you can even add a member to polish the design.

Then there's the grind of batch-producing materials. In swarm mode, you input a topic, audience, and style. The system assembles a team that splits the work—research, structure, content, integration—running in parallel instead of one after another. The claim is that a 200-page PPT can be delivered in 20 minutes. That's the kind of speed that changes how you plan your day.

Code, Music, and the Next Colleague

WorkSwarm isn't just for documents. There's a Code space with the same team logic. Want to recreate an arcade game you remember from a store that no longer exists? A swarm spins up a dev team: frontend and backend working on separate branches, merging to main, a tester checking things, a lead coordinating. Midway through, you get a new idea—just @mention a member to add a requirement. Or you want a calorie tracker. Describe it in one sentence, and the system plans the task, writes ArkTS code, compiles it, and produces an installable app for HarmonyOS PC.

Two demos, one song, one shared document. Neither is flashy, but together they show what WorkSwarm is aiming for: it can gather members, continue their work, use real tools, and always leave room for a human. The tagline says it plainly: One Platform, Super Teams. Complex work, handed to a team that's got each other's backs.

openJiuwen is also inviting developers to join its community, with free tokens for testing, a recognition program, and a path from newcomer to evangelist. If you're curious how a swarm of agents might handle your next project, the docs and downloads are open.