Huawei's open-source WorkSwarm brings swarm agents to HarmonyOS PC
The multi-agent trend crossed a new threshold this week: openJiuwen, the open-source agent platform built by Huawei's 2012 Labs, Huawei Cloud and the terminal team, launched WorkSwarm — a swarm of office agents that collaborate as a team, not a single assistant — and it's the first of its kind on the HarmonyOS PC app store.
Huawei-backed openJiuwen upgraded its JiuwenSwarm engine into WorkSwarm, a full office-agent suite that ships on the HarmonyOS PC app store with Windows and Mac support. The platform, released under Apache 2.0, frames this as the first engineering implementation of "Coordination Engineering" — the shift from making one agent smarter to making many agents work together. WorkSwarm spins up role-playing agents on demand: a reporting task assembles a research, drafting and review crew that passes shared context downstream, while a single-sentence coding request can spawn a frontend/backend team that builds, compiles and installs a working app. The demos are the fun part — seven agents (lyricist, composer, arranger, vocalist, and more) produced a full song with versioned revisions, and a 200-page deck shipped in 20 minutes via parallel task chains.
The technical pieces are what make this more than a demo reel: heterogeneous model routing assigns heavy-reasoning models to complex roles and lightweight ones to simple tasks, a layered memory system lets one swarm inherit another's experience, and leader and member agents can run on different devices — the PC hosts the leader while phones and tablets act as teammate nodes. That distributed architecture is clearly the HarmonyOS bet: it turns every Huawei device into an agent node, and it gives the Windows/Mac release the same collaboration model. openJiuwen also runs a Swarm Skills Hub where proven workflows are saved and reused, and OfficeClaw already builds on the engine for document automation.
The read: everyone is racing to ship agents that do real work; Huawei is betting the next differentiator is agents that hand off work to each other. It's early — these are orchestration demos until teams actually run on them daily — but open-sourcing the orchestration layer (rather than keeping it inside a closed office suite) is the interesting move, because it invites the community to build the skills that make the swarm useful.
Soup, an open-source tool, now fine-tunes an 8B model on a 4GB laptop GPU. The project, which surfaced on Hacker News and has been picked up by Product Hunt and international coverage, uses layer streaming — moving model layers between VRAM and system memory during training — to squeeze full fine-tunes into machines that previously couldn't load the weights, let alone train them. One YAML config drives the whole run. It's part of a broader wave of tools pushing fine-tuning off the server rack and onto consumer hardware, which matters for hobbyists and small teams who want to adapt models without renting GPUs; the tradeoff is slower training and tighter memory management. Worth a look if you've been priced out of fine-tuning.
What to watch: whether Huawei's swarm agents land beyond the HarmonyOS ecosystem — and whether "Coordination Engineering" becomes a category, or just a buzzword for a well-demoed group chat between models.
Do you want your AI assistants working as a team, or one very good assistant? Tell us in the comments.
Sources: openJiuwen (GitHub) · Soup (GitHub)
