Senin, 10 Agustus 2026

openJiuwen’s Agent Swarm: When AI Agents Finally Work as a Team

openJiuwen Agent Swarm

Over the past two years, the AI agent space has gone through wave after wave of "engineering paradigm shifts". From Prompt Engineering to Context Engineering, and on to Harness Engineering, each iteration has focused on making individual agents smarter and more reliable.

However, complex tasks in the real world are never completed by a single person.

For example, multi-disciplinary medical consultations require experts from various fields working in tandem.

Similarly, large-scale software engineering projects demand product managers, frontend engineers, backend developers, and QA testers pushing forward together.

And yet, AI industry is still trying to build an "all-powerful" single agent rather than a well-coordinated team of agents.

This is precisely the problem openJiuwen aims to address.

What is openJiuwen?

openJiuwen is an open-source AI agent platform jointly built by Huawei Laboratories, Huawei Cloud, and the Terminal Xiaoyi team, released under the Apache 2.0 license. Its goal is to help developers build production-grade AI agents and enable large-scale multi-agent collaboration.

In May 2026, the openJiuwen community released a major component: JiuwenSwarm. It marks the first engineering implementation of "Coordination Engineering", moving the focus from "making a single agent smarter" to "making multiple agents collaborate efficiently."

In collaboration with Huawei's Terminal Tablet & PC Product Line, Software Department, and other teams, openJiuwen launched the HarmonyOS PC version of JiuwenSwarm. This means JiuwenSwarm shares the exact same technical DNA with the HarmonyOS ecosystem right from the architectural foundation.

What makes JiuwenSwarm different?

At the core of JiuwenSwarm is the Agent Swarm mechanism, which enables multiple AI agents to autonomously divide labor, dynamically negotiate, and collaborate efficiently.

Below are its key features.

Heterogeneous model routing for different tasks

JiuwenSwarm supports routing members to different models, providing appropriate model capabilities for different roles, thus allocating heavy-reasoning models to complex logic and lightweight models to simple tasks. This reduces system load and improves overall execution quality. Developers simply declare the role-to-model mapping in a configuration file, and the system automatically manages scheduling.

JiuwenSwarm has already achieved seamless integration with platforms like HarmonyOS Xiaoyi and Lark (Feishu). Developers can build JiuwenSwarm-mode agents that are invokable directly via HarmonyOS devices.

Layered memory architecture

Paired with a layered persistent memory system, JiuwenSwarm retains full-dimensional records across sessions, including operation history, enterprise profiles, and policy guidelines. Through a cross-Swarm experience inheritance mechanism, the experience accumulated by one Swarm can be directly inherited by another.

Distributed deployment

JiuwenSwarm supports distributed Agent Swarms. The Leader Agent and swarm members can run across different processes, different nodes, or even entirely different servers. This capability is especially critical for tasks requiring access to isolated environments, such as intranet databases.

This architecture naturally adapts to the multi-device distributed scenarios of HarmonyOS PC. In the HarmonyOS PC "Agent PC" environment, the PC acts as the home base for the Leader Agent, while HarmonyOS devices like phones and tablets act as Teammate Agent nodes, working together without needing extra adaptation.

With the Channels mechanism supporting platform integrations like Xiaoyi out of the box, developers can build multi-agent applications for the HarmonyOS PC ecosystem straight away.

Observability and self-evolution

JiuwenSwarm provides an intuitive interactive interface, helping users track Swarm running states, task progress, and member workloads in real time.

Simultaneously, its experience closed-loop mechanism captures task breakdown paths, role allocation strategies, tool invocation sequences, and result evaluations, converting them into reusable "task templates." The next time a similar task arises, the Swarm can skip the planning phase and jump straight into execution.

In enterprise office scenarios, OfficeClaw has already leveraged JiuwenSwarm to achieve automated collaboration across content generation, document processing, and knowledge search.

Meanwhile, the Swarm Skills Hub has attracted contributions from numerous developers, fostering a virtuous ecosystem loop.

Where it's headed

In openJiuwen's commercialization roadmap, the native HarmonyOS agent system represents one of the largest and strategically most significant deployment cases.

With JiuwenSwarm now open source, developers can directly build multi-agent applications for the HarmonyOS PC ecosystem using this framework.

The project is officially open for the community on GitHub under the Apache 2.0 license.



from It's FOSS https://ift.tt/hvI4Hce
via IFTTT

Tidak ada komentar:

Posting Komentar