JiuwenSwarm: From Model Scale‑Up to Agent Scale‑Out

On July 24‑25, the openJiuwen community attended the DataFun "Agentic AI Summit – Super‑Intelligent Agent System Architecture" in Shenzhen, where Huawei AI foundation software scientist and openJiuwen technical committee chair Jin Xuefeng delivered the keynote titled "JiuwenSwarm: From Model Scale‑Up to Agent Scale‑Out".

The talk first reviewed the three recent paradigm shifts in agent architecture: Prompt Engineering, Context Engineering, and Harness Engineering. These paradigms mainly serve efficiency‑oriented, single‑agent tools with low failure cost and loose process boundaries.

When agents move into enterprise‑grade production systems, new challenges arise: tasks span multiple teams and systems, require long‑running collaborative workflows, and demand a multi‑agent coordination system. Enterprises also have mission‑critical scenarios that cannot rely on the existing Agent Loop model.

To address these challenges, JiuwenSwarm proposes two new engineering concepts—Coordination Engineering and Symbiosis Engineering—and builds an AgentOS that provides the necessary capabilities for large‑scale, multi‑agent collaboration in production environments.

The openJiuwen platform, co‑developed with Huawei, was officially open‑sourced in December 2025. It offers core capabilities such as Coordination Engineering, full‑stack agent self‑evolution, and compute‑infrastructure affinity, and has spawned representative AI agents like JiuwenSwarm, Jiuwen Symbiosis, CodingAgent, and DeepSearch, establishing a new swarm‑intelligence architecture paradigm.

To date, the community has amassed over 33 K stars on GitHub and deployed openJiuwen in more than 30 industry partners across finance, research, manufacturing, and other sectors, bringing agent capabilities into deep‑water enterprise applications. Huawei Cloud’s AgentArts has also integrated openJiuwen into its commercial platform for out‑of‑the‑box use.