Swarm Agent is here! The openJiuwen Community releases JiuwenSwarm, pioneering a new paradigm for Coordination Engineering

TL;DR

• The open-source AI Agent platform community openJiuwen, backed by Huawei, released JiuwenSwarm, a multi-agent swarm system designed to solve coordination engineering challenges.
• JiuwenSwarm achieved a State-of-the-Art (SOTA) score of 94.2% on the PinchBench evaluation, outperforming OpenClaw (91.6%) while reducing average token consumption by 34.8%.
• The system introduces a four-component architecture: Agent Swarm for collaboration, Swarm Skills for experience retention, Swarm Skills Hub for sharing, and an auto-evolution engine for continuous improvement.
• It supports two human-AI interaction modes: HOTS (Human on the Swarm) for command and control, and HITS (Human in the Swarm) for immersive participation as a team member.
• The framework demonstrates practical applications in medical diagnostics with a simulated team of 23 specialists and in Ascend operator development, highlighting improvements over single-agent performance.

Why It Matters

This release is significant because it shifts the focus of AI engineering from optimizing single agents (Harness Engineering) to orchestrating teams (Coordination Engineering), addressing the complexity of real-world tasks that require multi-role collaboration. For practitioners, it provides a fully open-source, end-to-end framework that standardizes how to build, share, and evolve multi-agent workflows, lowering the barrier to entry for deploying complex autonomous systems in domains like healthcare, education, and software development.

Key Data

• Benchmark Performance: JiuwenSwarm achieved a 94.2% composite score on PinchBench, compared to 91.6% for OpenClaw, representing a nearly 3-point improvement.
• Cost Efficiency: The system demonstrated a 34.8% reduction in average token consumption compared to the baseline, combining higher accuracy with lower cost.
• Memory Capability: On the LOCOMO long-term dialogue evaluation, openJiuwen achieved an 85% memory accuracy rate using an 8B large language model for processing, retrieval, and discrimination.
• Team Scale Example: A medical collaboration case study utilized a team of 23 distinct AI medical specialists to perform dynamic triage and joint diagnosis.

Technical Details

• Coordination Engineering Paradigm: The system extends the lifecycle from Prompt to Context to Harness Engineering, introducing Coordination Engineering to handle task decomposition, role assignment, and failure feedback loops among multiple agents.
• Agent Swarm & Routing: The core mechanism allows multiple agents to autonomously divide labor and negotiate dynamically. It supports model routing, assigning different LLMs to specific roles based on capability requirements to optimize load and performance.
• Swarm Skills & Auto-Evolution: A key feature is the "Swarm Skills" module, which standardizes team best practices into reusable assets. The evolution engine monitors execution trajectories to automatically generate new skills or update existing ones (e.g., adding new roles or constraints) upon user approval, creating a flywheel of continuous improvement.
• OpenJiuwen Harness: The underlying execution layer utilizes a DeepAgent architecture with optimized context engineering and long-term memory mechanisms, ensuring that individual agents within the swarm maintain high task execution stability and error recovery capabilities.

Industry Insight

The industry is moving from "super-individual" agent optimization to "team-based" organizational intelligence; enterprises should begin exploring frameworks that allow for the composition of specialized agents rather than relying on single generalist models for complex, long-horizon tasks. The introduction of a "Skills Hub" suggests a future market for multi-agent workflows where verified collaboration patterns become tradable, standardized assets, similar to how code libraries have evolved for single functions. The significant reduction in token cost alongside improved accuracy suggests that coordinated, role-specific agents are more efficient than single, overloaded agents, which has direct implications for the economic viability of deploying advanced AI agents at scale.

zation, dynamic negotiation, and collaborative evolution of multiple agents working together to achieve complex goals.

Q: Is JiuwenSwarm available for immediate use?
A: Yes, the platform is fully open-source. Developers can access the code and skill repositories on GitHub and AtomGit to deploy their own multi-agent swarms and contribute to the Swarm Skills Hub.