openJiuwen Open Sources Collective Intelligence Agent "JiuwenSwarm"

openJiuwen Unveils Collective Intelligence Agent Framework “JiuwenSwarm”

On May 18, 2026, the openJiuwen community, an open-source AI agent platform supported by Huawei, announced the release of “JiuwenSwarm,” a collective intelligence agent framework, as open source. This marks a paradigm shift from a single, high-performance agent (akin to a “lone lobster”) to a swarm of cooperative agents (like a “colony of bees”) working together autonomously.

Toward the Implementation of “Coordination Engineering”

According to openJiuwen, the paradigm of AI agent development has evolved from “prompt engineering” to “context engineering” and then to “harness engineering.” While harness engineering aimed to maximize the performance and reliability of individual agents, the next challenge lies in enabling multiple agents to collaborate like an elite team. Complex real-world tasks—such as large-scale software development or medical diagnosis—are often achieved by teams with diverse expertise. This philosophy, termed “Coordination Engineering,” is embodied in JiuwenSwarm as its practical implementation.

Core Technological Framework of JiuwenSwarm

JiuwenSwarm is built on four key components designed to realize collective intelligence:

Agent Swarm: This is the core of the system, providing mechanisms for multiple agents to autonomously divide tasks, negotiate dynamically, and collaborate efficiently. It supports routing to different models and assigns the most suitable model to each role, enhancing overall efficiency.

Swarm Skills: Best practices, standard operating procedures (SOPs), and role configurations for team collaboration are packaged as reusable “team-level skills.” These turn the collaboration patterns of successful agent teams into assets that can be immediately utilized.

Swarm Skills Hub: This platform enables the sharing, distribution, reuse, and derivative creation of accumulated Swarm Skills within an open ecosystem. It facilitates the circulation of collaborative knowledge among developer communities.

Self-Evolving Swarm Skills: During system operation, JiuwenSwarm continuously observes task execution traces, automatically learning and extracting effective new collaboration patterns. With user approval, these are registered as new Swarm Skills, creating a “flywheel effect” where the system strengthens the more it is used.

Human Collaboration: HOTS and HITS Modes

In addition to agent-to-agent collaboration, JiuwenSwarm is designed to allow humans to participate in the swarm. In HOTS (Human on the Swarm) mode, humans act as commanders of the agent team, monitoring task progress and bottlenecks, and intervening or adjusting as needed. In contrast, HITS (Human in the Swarm) mode integrates humans as equal members of the team, allowing for real-time collaboration and an immersive cooperative workflow.

Practical Demonstration: Efficiency in Ascend Operator Development

During the announcement, the practical effectiveness of JiuwenSwarm was demonstrated using Huawei’s Ascend AI processor operator development scenario. By employing JiuwenSwarm’s coding-oriented UI mode, the roles of algorithm design, kernel implementation, and performance optimization were distributed among multiple agents. The agents collaboratively translated algorithms from academic papers into engineering implementations. Compared to single-agent generation, JiuwenSwarm significantly improved the efficiency and quality of developing complex operators.

This announcement suggests a shift in focus within AI agent development from individual capabilities to coordinated collaboration. The success of JiuwenSwarm as an open-source project will depend on how extensively it can attract community participation and achieve practical applications.