AI News HubLIVE
In-site rewrite1 min read

openJiuwen Community Unveils JiuwenSwarm: Ushering in the Era of Swarm Intelligence

The Huawei-backed openJiuwen community has launched and open-sourced JiuwenSwarm, a swarm intelligence platform that enables multiple AI agents to collaborate efficiently and autonomously evolve. It introduces the Coordination Engineering paradigm, marking a shift from single-agent to multi-agent systems.

Source量子位Author: 思邈

The openJiuwen community, an open-source AI agent platform backed by Huawei, has announced the release and open-sourcing of JiuwenSwarm, a swarm intelligence system designed to orchestrate multiple AI agents into a cohesive, self-evolving team. This marks a paradigm shift from single-agent harness engineering to Coordination Engineering, a new framework for multi-agent collaboration.

JiuwenSwarm comprises four key components: Agent Swarm, which provides the underlying mechanism for agents to dynamically divide tasks and negotiate; Swarm Skills, which encapsulates best practices and standard operating procedures into reusable team-level capabilities; Swarm Skills Hub, an open marketplace for sharing and remixing these skills; and a self-evolution engine that continuously improves team and individual agent performance based on execution trajectories.

A distinctive feature is the human-agent collaboration modes: HOTS (Human on the Swarm) allows humans to act as commanders, monitoring and intervening at a high level, while HITS (Human in the Swarm) lets humans participate as equal team members, collaborating with agents in real-time. Practical demonstrations include multi-agent coding for Ascend operator development, multidisciplinary medical consultations with 23 AI specialists, short video production that improves with use, and immersive games like Werewolf where humans play alongside AI.

Performance benchmarks underscore the platform's strength. On the PinchBench agent evaluation suite, JiuwenSwarm achieved a state-of-the-art accuracy of 94.2%, outperforming OpenClaw's 91.6% while consuming 34.8% fewer tokens. Its memory system also scored 85% accuracy on the LOCOMO long-term dialogue benchmark using an 8B model.

By open-sourcing the entire stack, openJiuwen invites developers to contribute swarm skills and build their own agent colonies. JiuwenSwarm is now available on AtomGit and GitHub, with the Swarm Skills Hub live at swarm skills.openjiuwen.com. The community aims to democratize swarm intelligence, enabling anyone to become an "AI beekeeper" in the age of collaborative agents.