JiuwenSwarm

JiuwenSwarm is a multi-agent system built on the openJiuwen framework for automating complex tasks from natural-language requests. A Leader agent decomposes a task, assembles a team of specialized agents that negotiate the work between them, and drives tool calls through to a result. It runs on one machine or across a cluster, and is reachable from a web workbench, a terminal UI, a desktop app, and everyday chat apps.

Conversations run in Agent mode, where a single agent plans and adjusts on its own, or Cluster mode (the default), where the Leader orchestrates several specialists. The workbench has a Work space for office and general tasks and a Code space for viewing and modifying a project directory. Swarmflow adds deterministic multi-stage workflows defined in Python, with human-in-the-loop steps, a team token budget, and run-tree monitoring in the terminal UI.

Skills improve over time: the system watches for error signals and user dissatisfaction and rewrites Skill definitions in response, and an Auto Harness loop uses evaluations to tune the harness itself without retraining model weights. Skills can be shared and installed through the Swarm Skills Hub.

Models come from Huawei Cloud MaaS, OpenAI, DeepSeek, DashScope, SiliconFlow, OpenRouter, other OpenAI-compatible APIs, or local deployments. Tools ask for approval before running by default, file access goes through a whitelist, and sensitive operations are intercepted.