Scaling Enterprise Multi‑Agent AI: Insights from the QunXia AI Salon

Recently, the QunXia Intelligent AI Native Application Open‑Source Developer Salon in Beijing concluded, drawing over 110 technical participants. The event shared practical experiences on enterprise‑scale multi‑agent AI deployment.

Agenda 1 – HiClaw: Enterprise‑Level Multi‑Agent Collaboration & Harness Engineering Best Practices

HiClaw is a multi‑agent collaboration platform for enterprise AI applications. Its core follows a Manager‑Team‑TL‑Workers architecture: the Manager breaks down and schedules tasks, the Team Leader coordinates sub‑tasks, and Workers execute specific skills. Communication relies on the Matrix protocol to make interactions between humans and agents, as well as between agents, transparent. The platform introduces an AI gateway for unified authentication, sandboxed execution environments for isolation, and a lightweight communication mechanism to address security credential isolation, SubAgent efficiency, and resource consumption. A live demo demonstrated building an AI team from scratch.

Agenda 2 – QwenPaw: Instantly Load Your Personal Intelligent Companion

QwenPaw is a personal AI assistant workstation built on the AgentScope‑AI ecosystem. It emphasizes local deployment, security, controllability, and efficient interaction. The system is optimized for small models, integrates a skill marketplace, tool‑call protection, and long‑term memory management, and supports various open‑source or private model back‑ends. Users can deploy the full agent workflow locally, preventing sensitive data leakage. QwenPaw provides a visual interface and debugging tools to help developers quickly construct, test, and iterate personal AI assistants for knowledge management, script automation, and office assistance while preserving privacy.

Agenda 3 – AgentScope‑Java Data‑Flywheel Q&A Agent Practice

The presented Q&A agent abandons traditional knowledge bases and lets LLMs read source code directly, treating "code as knowledge" to avoid stale information and high maintenance costs. The system adopts a four‑layer architecture: an Access layer handles user requests, a Skill layer encapsulates inference logic, a Tool layer invokes code‑search and execution capabilities, and a Data layer supplies contextual support. Through an AgentLoop mechanism, the solution continuously observes, auto‑evaluates, performs regression testing, and optimizes policies, forming a closed‑loop iteration. In internal technical Q&A scenarios, this approach markedly improved answer accuracy and response quality, demonstrating the feasibility and engineering value of source‑code‑driven AI assistants.

Agenda 4 – Nacos Skill/Worker Registry: Enterprise‑Private AI Registry for Rapid Standardization

Nacos introduces an enterprise‑private AI registry to tackle security, permission, stability, and governance challenges of Skills and Workers in production. The registry embeds a security‑scan mechanism, supports multi‑version gray releases, namespace isolation, and full‑link operation auditing, ensuring controlled AI capability rollout. A Worker is defined as a reusable unit containing a complete workflow, configuration, SOP, and dependencies, allowing teams to invoke verified templates and avoid duplicate development. This extends traditional micro‑service governance concepts to the AI domain, providing foundational infrastructure for standardized, traceable, and highly available AI asset management.

Hands‑On Session

The event concluded with a hands‑on workshop where instructors demonstrated HiClaw deployment and simple scenario experiences, guiding participants through interactive exercises and fostering lively discussion.