Self-Adaptive Digital Team
A fully visible, self-evolving ReAct Digital Team that intelligently decides when to be fast and when to go deep.
MultiAgentSwarm v4.1.0 不再是简单的“多个LLM并行聊天”,而是一个完全可视化、智能决策、自进化 ReAct 数字团队 —— 它完美还原经典 ReAct 架构,同时注入群体智能、动态规划、分层终身记忆与生产级交付能力。
MultiAgentSwarm v4.1.0
A fully visible, self-evolving ReAct Digital Team that intelligently decides when to be fast and when to go deep.
We rejected the “use everything on every task” anti-pattern.
Instead, we built a swarm that knows itself — automatically choosing the right level of intelligence for each request.
This is the only swarm that:
- Intelligently routes every task: Simple / Medium / Balanced / Complex
- Forces visible ReAct thinking in every single agent response
- Delivers Balanced mode as the default sweet spot (3 agents, planning + tools, best quality/speed ratio)
- Runs Tree-of-Thoughts with 3 concurrent exploration branches
- Spawns Dynamic Agents on-the-fly with true parallel subtask execution (Hierarchical Supervisor)
- Maintains lifelong hierarchical memory (PrimalMemory + Vector + Knowledge Graph) with automatic decay & distillation
- Performs all self-evolution completely in the background (Auto-Eval + Active Distillation + memory persistence) — zero freezing after answer delivery
- Generates & validates Master Plan with Benjamin review pass + dynamic refresh
- Supports instant task cancellation, streaming output, smart image compression, and tool overrun friendly fallback
- Ships production-ready: WebUI + Feishu long-polling + Email integration + smart file delivery
A local swarm that feels like a real team:
Fast when you need speed. Deep when you need excellence. Never makes you wait.
| Feature | MultiAgentSwarm v3.2.0 | Others (AutoGen/CrewAI/LangGraph) |
|---|---|---|
| Task Routing | ✅ Auto Simple/Medium/Balanced/Complex | ❌ Always full swarm |
| ReAct Visibility | ✅ Forced Thinking/Action in every response | ❌ Opaque or optional |
| Balanced Mode | ✅ Fixed 3 agents (Grok+Harper+Benjamin) + Master Plan + 1 light debate | ❌ No dedicated sweet-spot mode |
| Adversarial Debate | ✅ Pro/Con/Judge + Meta-Critic (quality-based) | ❌ Manual only |
| Master Plan + Benjamin Review | ✅ Auto-generated + validated every run | ❌ None |
| Memory System | ✅ PrimalMemory (tree+atomic+decay) + Vector + Knowledge Graph | ❌ Ephemeral |
| WebUI | ✅ Real-time streaming + cancel + expandable ReAct panel + file download | ❌ Requires extra work |
| Feishu Integration | ✅ Official long-connection + auto 👍 | ❌ Not supported |
| Token Efficiency | Balanced saves 60-80% vs Complex | ❌ Always expensive |
1. 🧭 Full ReAct Visualization (Core of v3.2.0)
- Every agent response must begin with:
Thinking:→Action:→Action Input: - Tool results clearly marked
【Observation】(red highlighted) - Real-time WebSocket streaming → you literally watch every agent think.
2. 📋 Dynamic Master Plan + Benjamin Validation
- Auto-generates structured plan
- Benjamin (Agent[2]) performs dedicated Plan Review Pass
- Refreshes plan when quality drops or every 3 rounds
3. 🟠 Balanced Mode — The Sweet Spot
- Fixed 3 agents: Grok (leader) + Harper (creative) + Benjamin (critic)
- Master Plan injection + Benjamin review + exactly 2 rounds
- Light adversarial debate only on round 2
- Quality ≈ Complex 80-85% at ~140% speed of Medium
4. Intelligent Routing (v4 Classifier)
- Rule + LLM hybrid
- Balanced is the default for analysis/reports/structured output
- Complex only when truly needed (real-time tracking + deep reflection) or user forces it (中英日 keywords supported)
5. Production WebUI & Integrations
- True per-agent streaming + expandable Thinking panel
- Upload images/PDFs → auto multi-modal or text parsing
- One-click Markdown export + smart file generation with
/uploadsdownload links - Feishu long-connection with instant 👍 reaction
- Task cancel button + heartbeat keep-alive
6. Memory & Knowledge Systems
- PrimalMemory (tree logs + atomic KB + exponential decay)
- VectorMemory (ChromaDB)
- Active Knowledge Graph + automatic distillation
| Mode | Agents | Rounds | Debate | Quality (vs Complex) | Typical Time | Best For |
|---|---|---|---|---|---|---|
| Simple | 1 | 1 | 0 | 60% | 1-3s | Greetings, quick facts |
| Medium | 2 | 1 | 0 | 75% | 8-15s | Concept explanation |
| Balanced | 3 | 2 | 1 | 80-85% | 18-28s | Analysis, reports, files |
| Complex | 4+ | 4-8 | 2-4 | 100% | 40-80s+ | Real-time deep tracking |
Balanced is now the recommended default for 70%+ of real tasks.
git clone https://github.com/yourname/MultiAgentSwarm.git
cd MultiAgentSwarm
uv pip install -r requirements.txtRecommended: Start with OpenSandbox (hard isolation)
WebUI (recommended): python webui.py → http://localhost:8060
Via EMAIL
- Configure email information in the configuration file.
- Send an email to the recipient's address (using the predefined subject).
openai:
default_model: "gpt-4o-mini"
context_limit_k: "128"
base_url: "http://localhost:11434/v1" # Ollama / vLLM / any compatible
advanced_features:
adversarial_debate:
enabled: true
trigger_strategy: "quality_based"
adaptive_reflection:
quality_threshold: 85
intelligent_routing:
enabled: true
force_complexity: null # "simple" / "medium" / balanced / "complex"All features are hot-reloadable via WebUI /api/config.
- Beautiful WebUI with real-time streaming + cancel button + expandable ReAct panel
- Upload images/PDFs/TXT/MD → auto multi-modal or text parsing
- Talk directly to Feishu (group/private) → bot replies automatically with 👍
- Generate downloadable reports with one click
- Export any conversation as Markdown
- 25+ Skills + instant skill extension via
/skills/ - Persistent PrimalMemory + Vector DB across restarts
- Full local model support (Ollama/vLLM or any OpenAI-compatible endpoint)
- Email support
MultiAgentSwarm/
├── webui.py # FastAPI + WebSocket + Feishu (main entry) + Email
├── multi_agent_swarm_v3.py # Core swarm logic
├── skills/ # 25+ ready skills + custom .py/.md
├── uploads/ # Uploads + generated files (directly downloadable)
├── static/index.html # Frontend
├── requirements.txt # Full dependencies
├── swarm_config.yaml # All settings
└── memory/ # PrimalMemory + Vector DB + Knowledge Graph
Python 3.10+
uv pip install -r requirements.txt # or pipNo GPU required. Model caching automatic. OpenSandbox optional (auto fallback).
Star ⭐ if you like the direction — it keeps us motivated!
MultiAgentSwarm WebUI
一个真正看得见Agent思考的自适应数字团队
MultiAgentSwarm v4.1.0 不再是简单的“多个LLM并行聊天”,而是一个完全可视化、智能决策、自进化 ReAct 数字团队 —— 它完美还原经典 ReAct 架构,同时注入群体智能、动态规划、分层终身记忆与生产级交付能力。
市面上的多智能体框架大多采用同一种粗暴思路:
“不管什么任务都把所有Agent扔上去 → 大量烧token → 祈祷出好结果”
我们从第一性原理出发,基于两个核心公理重新构建:
- 真正的智能不是用更多Agent,而是知道什么时候该快、什么时候该深、该并行什么、该记住什么。
- 高质量绝不能以让用户等待为代价 —— 答案必须即时呈现,所有自进化过程必须完全后台异步完成。
成果:MultiAgentSwarm v4.1.0 —— 目前唯一同时做到以下全部特性的本地多智能体系统:
- 智能任务路由:自动判断并切换 Simple / Medium / Balanced / Complex 四种模式
- Balanced 甜点模式:默认最优选择(规划 + 工具 + 轻量反思,质量接近 Complex,速度接近 Medium)
- Tree-of-Thoughts 多分支并行探索:3条思考路径同时推进
- 动态Agent工厂 + Hierarchical Supervisor:真正并行执行子任务(按需生成专家并发工作)
- 强制可见 ReAct 思考:每个Agent都必须先输出 Thinking → Action(完全透明)
- 分层终身记忆系统:PrimalMemory + Vector + Knowledge Graph + 自动衰退与蒸馏
- 完全异步后台自进化:答案显示后立即可用,记忆保存、Auto-Eval、Active Distillation 全在后台运行(零冻结)
- Master Plan 生成 + Benjamin 验证:动态规划 + 严格审查 + 实时刷新
- 生产级集成:开箱即用的 WebUI + 飞书长连接 + 邮件智能回复 + 智能文件生成与下载
一句话总结:
这是一个看得见思考、懂得节制、持续进化的本地数字团队 —— 既能秒回简单问题,也能高质量完成复杂报告,还永远不会让你等后台处理。
| 特性 | MultiAgentSwarm v3 | AutoGen / CrewAI / LangGraph |
|---|---|---|
| 智能路由 | ✅ 自动简单/中等/复杂 | ❌ 永远全量群聊 |
| 对抗式辩论 | ✅ Pro/Con/Judge + 元批评 | ❌(需手动实现) |
| Primal终身记忆 | ✅ 树状日志 + 原子KB + 衰退 | ❌ 临时记忆 |
| 知识图谱 + 蒸馏 | ✅ 实时概念图谱 | ❌ 无 |
| 自适应反思深度 | ✅ 质量达标自动停止 | ❌ 固定轮数 |
| 强制ReAct透明格式 | ✅ 每轮都显示思考过程 | ❌ 黑箱 |
| WebUI + 流式 + 取消按钮 | ✅ 开箱即用 | ❌ 需要额外开发 |
| 飞书企业级长连接 | ✅ 零配置 | ❌ 不支持 |
| 智能文件生成与下载 | ✅ 自动识别意图并生成可下载报告 | ❌ 手动 |
| Token & 时间节省 | 简单任务节省60-80% | ❌ 永远昂贵 |
| 本地模型支持 | ✅ 任何兼容OpenAI接口的LLM(Ollama、vLLM、DeepSeek、Qwen等) | ❌ 支持有限或需额外配置 |
核心升级: 核心特性:
- 智能任务路由(Simple / Medium / Balanced / Complex)+ Balanced甜点模式
- Tree-of-Thoughts 多分支并行探索(3 Agent 并发)
- 动态 Agent 工厂 + Hierarchical Supervisor 真正并行子任务执行
- 分层记忆系统(PrimalMemory + Vector + Knowledge Graph)
- 完全异步后台处理(记忆保存 + Auto-Eval + Active Distillation + 无任何冻结)
- Master Plan 生成 + Benjamin 验证 Pass + 动态刷新
- 任务取消支持 + 流式输出 + 图片智能压缩 + 工具超限友好兜底
- Balanced 模式严格锁定 3 个 Agent(Grok + Harper + Benjamin)——质量与速度最佳平衡点
- 所有 Agent 强制输出 ReAct 三段式思考过程(实时可见)
- WebUI 完整流式 + 文件智能生成下载 + 飞书长连接 👍
1. 🧭 显式 ReAct 思考过程(架构图 100% 对齐)
- 每条 Agent 回复必须以以下格式开头:
Thinking:(原因分析)
Action:(调用工具名称或 Final Answer)
Action Input:(参数 JSON 或最终答案摘要) - 工具返回结果独立标记为 【Observation】(红色醒目)
- WebSocket 实时流式输出,用户和开发者可完整看到思考链路。
2. 📋 动态 Master Plan 刷新(动态规划闭环)
- 每 3 轮或质量 < 75 分时自动刷新 Master Plan
- 完美闭合架构图“更新prompt”循环
- 所有 Agent 始终对齐最新规划。
3. 🧭 Intelligent Routing(智能任务路由) ★ 旗舰特性
- 自动判断 Simple / Medium / Complex
- 规则 + LLM 双重判断 + 自动降级
4. 🥊 Adversarial Debate + Meta-Critic
- Pro / Con / Judge 三角色并行辩论,每轮强制先挑刺
- Meta-Critic 二次综合评估
5. 🏭 Dynamic Task Decomposition + 🧠 Active Knowledge Graph + PrimalMemory
- 自动拆解 4-7 个子任务并智能分配
- 实时实体-关系提取 + 重要性蒸馏 + 树状日志 + 原子 KB + 衰退机制
6. 📈 Adaptive Reflection Depth(自适应反思深度)
- 质量 ≥85 分立即停止
- 质量收敛(Δ<3)自动停止
- 全部参数通过 API 实时可调
7. 🌐 美观生产级 WebUI(v3.2.0 增强版) ★ 全新
- 真实逐 Agent WebSocket 流式输出 + 可展开「🤔 思考过程」面板
- Master Plan 动态刷新实时日志可见
- 多会话管理 + 一键导出 Markdown
- 文件上传(PDF/图片/文本,最大10MB)+ 中文文件名自动净化
/uploads静态挂载 → 修改后的报告/Excel/PDF 可直接点击下载- 任务取消按钮 + 30秒心跳保活
- 完整飞书官方 SDK 长连接 + 收到消息立即自动👍反应
- 邮件交互
8. 🔒 OpenSandbox 双模式代码执行器
- 已安装 → Docker 硬隔离(推荐)
- 未安装 → 自动回退 + 醒目安装提示
9. 🌍 本地模型支持 + 25+ 开箱即用 Skill
- 任何兼容 OpenAI 接口的 LLM(Ollama、vLLM、DeepSeek、Qwen 等)
- 25+ Skill 开箱即用 + 极易扩展:把
.py或.md扔进/skills/即可立即加载
| 指标 | v2.9.2 | v3.1.0 | v3.2.0(现在) | 提升幅度 |
|---|---|---|---|---|
| 简单任务耗时 | 8-12s | 1-3s | 1-3s | -75% |
| 复杂任务质量 | 8.0/10 | 9.5/10 | 9.7/10 | +21% |
| 思考过程透明度 | 无 | 部分 | 完整实时可见 | 革命性提升 |
| 规划漂移(5+轮) | 中 | 低 | 几乎为 0 | 彻底解决 |
| Token 消耗 | 基准 | -40~60% | -45~65% | 进一步节省 |
| 文件处理 | 无 | 基础 | 中文名净化 + 下载支持 | 生产级可用 |
git clone https://github.com/yourname/MultiAgentSwarm.git
cd MultiAgentSwarm
uv pip install -r requirements.txt启动 OpenSandbox(推荐)
opensandbox-server init-config ~/.sandbox.toml --example docker
opensandbox-server启动 WebUI
python webui.py通过 EMAIL
- 在配置文件中配置邮箱信息
- 发送邮件到收信邮件地址(使用设定好的标题)
CLI 测试
python multi_agent_swarm_v4.py享受构建属于你自己的完全透明数字团队吧! 🚀
最后更新:2026 年 3 月 3 日
版本:v3.2.0
