JiuwenSwarm Launches Coordination Engineering for the ‘Beekeeping’ Era of AI Agents

From Harness to Coordination

openJiuwen, the Huawei‑backed open‑source AI Agent platform, announces JiuwenSwarm, a new swarm‑intelligence framework that marks the transition from Harness Engineering to Coordination Engineering. The article traces the evolution of AI‑Agent engineering: Prompt Engineering → Context Engineering → Harness Engineering → Coordination Engineering, positioning the latter as the engineering paradigm for multi‑Agent collaboration.

Coordination Engineering Core Design

The shift raises four inter‑linked questions: how agents autonomously divide work and negotiate, how best‑practice collaboration is captured as reusable assets, how these assets circulate among developers, and how the system stays dynamic rather than static. JiuwenSwarm answers with a full‑stack technical suite:

Agent Swarm – the core that enables multiple agents to self‑assign tasks, dynamically negotiate, and cooperate efficiently. It supports routing different agents to models best suited for specific roles, reducing load and improving overall performance.

Swarm Skills – a library that standardises team‑level best practices, SOPs, role configurations and scheduling strategies into reusable "team skills" that can be plugged into any scenario.

Swarm Skills Hub – an open marketplace where these skills are shared, reused and iteratively improved by the developer community.

Swarm Skills Self‑Evolution – a flywheel that observes task execution traces, automatically extracts reusable skills, and updates them after user approval, strengthening both the team and individual member capabilities over time.

Human Interaction Modes

JiuwenSwarm defines two collaboration modes for humans:

HOTS (Human on the Swarm) – the human acts as a commander, observing the whole swarm, adjusting priorities, swapping agent roles, or issuing fine‑grained commands.

HITS (Human in the Swarm) – the human becomes a participant agent, working side‑by‑side with AI teammates in real‑time, similar to a player in a Werewolf game.

Real‑World Case Studies

Case 1: Multi‑disciplinary Medical Consultation – a team of 23 AI medical specialists dynamically forms a joint diagnosis, achieving broader coverage and higher accuracy than a single expert.

Case 2: Multi‑Agent Operator Development – in Ascend operator generation, separate agents handle algorithm design, kernel implementation and performance optimisation, improving development efficiency and quality.

Case 3: Short‑Video Production – the swarm creates a video, the evolution engine extracts a reusable skill, and subsequent productions automatically improve titles, styles and platform adaptation, illustrating the "use more, get stronger" effect.

Case 4: Werewolf Game with Multiple Models – different AI models assume distinct player roles; humans can switch between HOTS (global control) and HITS (immersive participation), demonstrating flexible human‑AI interaction.

Case 5: Immersive Multi‑Disciplinary Tutoring – students and parents interact with AI teachers across subjects, receiving personalised guidance and progress reports.

Benchmark Performance

On the PinchBench benchmark (covering code development, creative writing, document handling, meeting management, content conversion, file operations), JiuwenSwarm scores 94.2% overall, surpassing OpenClaw’s 91.6% by nearly 3 points. It also reduces average token consumption by 34.8%.

In the LOCOMO long‑term memory benchmark, the openJiuwen Harness achieves 85% memory accuracy with an 8B model, outperforming major industry memory systems.

Underlying Harness and Open‑Source Commitment

The strong performance stems from the openJiuwen Harness, built on DeepAgent architecture, context engineering and long‑term memory mechanisms. All components—Agent Swarm, Swarm Skills, Hub and self‑evolution—are fully open‑source, with repositories on AtomGit and GitHub, inviting community contributions.

Conclusion

JiuwenSwarm demonstrates that coordinated multi‑Agent systems can be delivered as a runnable, installable, co‑creatable, fully open‑source stack, positioning the community at the forefront of the AI "beekeeping" era.