openJiuwen-ai/jiuwenswarm

JiuwenSwarm is a multi-agent orchestration system built around a Leader-Teammate model: a Leader agent decomposes a complex task, assembles a team of specialized agents, and those agents negotiate dynamically to get the work done. That architecture is not novel in itself, but the system layers several concrete mechanisms on top of it that are worth examining.

The most interesting piece is Skill Self-Evolution. The system watches for error signals and user dissatisfaction, then automatically rewrites Skill definitions in response — no model-weight training involved. The README calls this an "Auto Harness" that improves through evaluation-driven optimization in practice. Whether that loop is robust in adversarial or ambiguous cases is not addressed, but the design intent is clear: capabilities should improve through use rather than through manual prompt engineering.

Swarmflow is the deterministic counterpart to that adaptive layer. It lets you define multi-stage workflows as Python scripts, with the Leader handing off between stage agents in a fixed sequence. It supports human-in-the-loop checkpoints via human and human_session step types, team-level token budgets, and a TUI run-tree monitor. That combination — scripted determinism plus optional human gates plus cost controls — is a practical answer to the reliability problem that pure autonomous orchestration creates.

The channel integration is broader than most comparable tools. Beyond the expected Telegram and Slack, the system supports Feishu, DingTalk, WeCom, Personal WeChat, WhatsApp, and Xiaoyi. Feishu and WeCom additionally expose a Digital Avatar feature. Capabilities vary by platform — some are private-chat only, others support group mentions — which reflects real deployment constraints rather than a uniform abstraction.

Distribution is handled at the process and machine level: Leader and Teammate agents can run across separate machines and coordinate at scale. The README does not specify the transport or consensus mechanism behind that, so the actual failure modes of distributed coordination are opaque from this source alone.

Security posture is explicit: every tool requires approval before execution by default, file access goes through a whitelist, and sensitive operations are intercepted. That default-deny stance is the right starting point for anything running shell-level tools, though the approval workflow can be relaxed per policy.

The project is in active beta — the most recent changelog entry, dated 2026-08-06, targets cold-start latency in the Agent instance launch path. Earlier entries added browser sub-agent isolation, multimodal Skill creation from visual input, and a code-diff display in the Code workspace. The version numbering (v0.2.4.beta3) signals that the API surface is still moving.

Installation is straightforward: pip install jiuwenswarm, then jiuwenswarm-init and jiuwenswarm-start, with a web frontend at localhost:5173. A separate TUI package is available. Desktop installers exist for Windows, macOS, and HarmonyOS — the last of which is an unusual target that signals the project's primary market.