openJiuwen Community: Opened New Recruitment for Swarm AI, Launching JiuwenSwarm to Start the 'Beekeeping' of Collective Intelligence

openJiuwen Community New Recruitment:重磅发布 JiuwenSwarm, Launching the "Beekeeping" Dawn

https://www.qbitai.com/2026/05/419515.html

2026-05-18 18:26:46 来源：量子位

The Bottom Layer Changed

允中 发自 凹非寺

量子位 | 公众号 QbitAI

Just recently, Huawei-backed open-Jiuwen released and made public the swarm-based AI Agent platform JiuwenSwarm, marking a significant step in the journey toward group intelligence.

This is not just a change in name—it's a shift in the underlying framework.

Multiple AI agents can work together like bees to achieve efficient collaboration and autonomous evolution, officially activating the "group intelligence" acceleration key, opening the way for the "beekeeping" era of AI.

This is behind open-Jiuwen's next major paradigm shift — Coordination Engineering (CoE) — which has been fully implemented.

To understand this upgrade, you must first answer the question:

Why is it now, from Harness to Coordination?

Why from Harness to Coordination?

Looking at the timeline more closely, from Prompt Engineering to Context Engineering, and then to Harness Engineering, the engineering paradigms in AI Agent development have been continuously evolving:

• Prompt Engineering: Debugging prompts to let models understand tasks.
• Context Engineering: Organizing agents' context, memory, tools, and state.
• Harness Engineering: The keyword term that has swept the industry since this year, covering agent engineering, trajectory management, error recovery, and long-range execution all the way to its peak.

Following this, the next engineering challenge is:

How can multiple agents work as a team?

Because real-world complex tasks such as cross-domain deep research, large software project delivery, multi-role decision-making, and complex business process orchestration — are never just one person's job, they require a team.

Software needs product + development + testing + SRE; education needs multi-disciplinary teachers + parents + oneself; healthcare needs a doctor + multiple medical experts — these are all examples of teams working together.

This is exactly what open-Jiuwen proposes as the next major paradigm shift: Coordination Engineering (CoE) — an engineering approach centered around "multi-agent collaboration."

And this time, open-Jiuwen has turned "cooperation" into a complete open-source engineering solution:

JiuwenSwarm.

Core Design Principles of Coordination Engineering

To truly enable a team of agents to work, several problems need to be solved:

• How do agents autonomously divide their roles and dynamically coordinate?
• What are the best practices and role combinations that can be standardized and reused?
• How can experience be shared and reused between developers, enabling continuous improvement?
• How does the system evolve over time, rather than just running endlessly?

These four questions are interlinked, each being a necessary extension of the previous.

JiuwenSwarm provides a comprehensive full-stack technology solution:

Agent Swarm, Swarm Skills, Swarm Skills Hub, and Swarm Skills Self-Improvement.

The four key components form a seamless chain:

Agent Swarm — Building a Multi-Agent Team

This is the core of the entire system.

Agent Swarm provides a mechanism for multi-agent collaboration, allowing them to autonomously divide roles, dynamically coordinate, and efficiently collaborate, transitioning from "single-person combat" to "a well-organized team."

JiuwenSwarm supports model routing for different agents, allowing for appropriate model capabilities based on roles, reducing load pressure, and enabling personalized teaching.

Swarm Skills — Transforming a Team into a Comprehensive Combat Capability

Agent Swarm addresses how to collaborate, while Swarm Skills addresses how to accumulate and standardize skills.

It encapsulates best practices, SOPs, role combinations, and scheduling strategies into standardized "team-level skills", enabling a "team with any scenario" combat capability.

Swarm Skills Hub — Sharing Team-Level Skills in a Shared Market

After skill accumulation, the next step is sharing.

Swarm Skills Hub creates an open ecosystem where team-level collaboration experiences are shared, reused, and reimagined.

https://swarmskills.openjiuwen.com/

Swarm Skills Self-Improvement — A Continuous Evolution of Power

The final ring of this loop is the most imaginative.

As teams execute tasks, JiuwenSwarm's self-improvement engine continuously observes task decomposition, role scheduling, and message exchanges — automatically identifying and replying to reusable Swarm Skills — and submits approval for user entry.

At the same time, it operates on two levels:

• Team Level: Automatically adjusts roles, adds constraints, and optimizes collaboration workflows based on task execution trajectories, enhancing leadership planning and control.
• Member Level: Collects and shares experience gained during real-world interactions, solving problems directly when encountering similar ones without repeating mistakes. Teams grow, and each member improves.

How Can Humans Participate in Collaboration: HOTS & HITS

From teamwork to experience沉淀, from skill sharing to continuous evolution, four core capabilities form a complete cycle.

However, on top of this framework, there's another fundamental yet practical question: How can humans work alongside this multi-agent team?

JiuwenSwarm offers two modes: HOTS (Human on the Swarm) and HITS (Human in the Swarm).

1. HOTS (Human on the Swarm): Human as the Leader

Humans stand at a higher position, observing the entire team's operation state: progress, role load, and collaboration bottlenecks.

When needed, humans can intervene immediately — adjusting task priorities, switching agent roles, or changing plans — with granular control over individual instructions, or broad control over a single direction.

2. HITS (Human in the Swarm): Human as a Member

Humans are no longer distant observers but part of the team, operating in the same environment, scene, flow, and collaboratively exploring.

They represent the "bee" in the swarm — a member of the team.

Like players in a game such as "wolf hunt," humans are immersed in this collaboration.

HITS is immersive participation, while HOTS is global coordination — these are the two most fundamental ways humans can work with the multi-agent team.

Practical Effect of JiuwenSwarm

Let's look at how JiuwenSwarm performs in various domains, giving you a firsthand sense of the power of Coordination Engineering.

Case 1: Multi-Agent Collaboration for Algorithm Development

JiuwenSwarm provides a TUI mode for coding, allowing experts to take on roles such as algorithm design, kernel implementation, and performance optimization — thus enabling the algorithm to move from paper to code.

The collaborative process is visible, with each expert taking their own role and collaborating to optimize the algorithm.

Swarm Skills Hub

Case 2: Multidisciplinary Medical Expert Teams for Diagnostics

A team of 23 AI medical experts collaborates dynamically, creating multiple expert members based on patient conditions. Each expert analyzes the condition and real-time communicates diagnostic results, seeking consensus and providing accurate diagnosis and recommendations.

The collaborative process is visible, with each expert performing their role and seeking consensus, leading to improved diagnostic quality.

Swarm Skills Hub

Skill Accumulation and Team Growth

Savvy team experiences can be accumulated through collaboration, leading to enhanced performance.

In case 3, users start a video creation task, and the JiuwenSwarm self-improvement engine identifies a reusable collaboration pattern, automatically generating a video creation Swarm Skill and submitting it for approval.

Using this skill again, the engine detects inconsistencies in the image style and the user's request for a video platform, generating new content with high click-through rate titles. This enhances the video effect further.

Swarm Skills Hub

Support for Different Models and Human Roles

JiuwenSwarm supports model routing and human roles (HOTS/HITS) configuration.

#### Case 1: Multi-model Participation in a Game

For example, in the wolf-hunt game, different agents use different models for their respective roles.

Additionally, humans can operate as "God's eye" observers, controlling the whole game globally — this is HOTS mode.

#### Case 2: Immersive Human Experience in a Game

Want to immerse yourself in multi-agent collaboration?

JiuwenSwarm offers HITS mode, where humans can be players in the game — they can be hunters, forecasters, or villagers — and interact with AI teammates.

Other agents listen to your messages, infer your identity, and decide whether to "take flight" or "vote" for you.

Swarm Skills Hub

Tips: Switching Between HOTS and HITS

To realize the freedom of switching between HOTS and HITS, you can refer to the following instruction:

Case 3: Immersive Multi-Disciplinary Course Guidance

Children and parents can "enter the scene" with other disciplines' intelligent agents, facilitating professional guidance.

When a human switches to a student identity, they interact with teachers, assessing students' understanding of the subject. When a human switches to a parent identity, they discuss supervision and motivation.

The "Academic Coach Team" skill is available on Swarm Skills Hub.

Behind the Coexistence: OpenClaw Provides Hard Power

JiuwenSwarm's swarm-based collaboration is very impressive, but each bee relies on openJiuwen Harness, which is also a strong foundation.

Without strong execution power from a single agent, even the most sophisticated collaboration mechanisms cannot be implemented.

This was verified by the authoritative benchmark set PinchBench, developed by Kilo.ai.

PinchBench evaluates Agent performance across a wide range of tasks, including code development, creative writing, document processing, meeting management, content conversion, file operations, etc.

Its task design is closely aligned with real-world business scenarios, and its evaluation dimensions are comprehensive, making it a reliable reference for measuring Agent execution ability.

• openJiuwen has an impressive memory mechanism, with a memory accuracy rate of 85% in the LOCOMO long-term dialogue benchmark (processed by an 8B large model), outperforming most industry-leading memory systems.
• These results are not random; they result from the excellent performance of openJiuwen Harness in DeepAgent architecture, context engineering, and long-term memory mechanisms, which have been refined over time to enable JiuwenSwarm's team members to achieve strong task execution capabilities.

Conclusion: Open Source, Together as "Bees in the Age of AI"

Looking back, from Harness Engineering to Coordination Engineering, and today’s JiuwenSwarm, openJiuwen has made a significant contribution to this field.

1. A coordination engineering philosophy

2. A full-stack technical system (Agent Swarm / Swarm Skills / Swarm Skills Hub / Self-evolving skills)

3. The JiuwenSwarm intelligent agent

Additionally, all features are fully open-sourced.

Multi-agent collaboration is a consensus, but it can be implemented as a complete, deployable, and open-source solution. Currently, only a few companies can achieve this.

AI Agent's vast horizon is not just about being omnipotent, but about building a group intelligence that collaborates, evolves, and adapts.

With JiuwenSwarm marking the first flag for this path, each user can easily build their own smart bee hive.

About openJiuwen

openJiuwen is an open-source AI Agent platform community supported by Huawei, built by the Huawei 2012 Lab and the Huawei Cloud AgentArts team. Its flagship intelligent agent, JiuwenSwarm, has fully developed expertise in key areas such as Harness Engineering, multi-agent collaboration, and self-evolution.

JiuwenSwarm (AtomGit): https://atomgit.com/openJiuwen/jiuwenswarm JiuwenSwarm (GitHub): https://github.com/openJiuwen-ai/jiuwenswarm Swarm Skills Hub: https://swarmskills.openjiuwen.com/

Please share your team skills with the Swarm Skills Hub, allowing the community to flow knowledge and experience.

Let us together cultivate bees to produce honey!

_All rights reserved. Unauthorized use or reproduction is prohibited. Violations will be strictly enforced._