{"slug": "maref-formal-governance-for-multi-agent-systems", "title": "Maref: Formal Governance for Multi-Agent Systems", "summary": "Maref, a multi-agent recursive engineering framework, launched as the first product focused on agent governance, featuring a 10-state Gray code state machine, TLA+ formal verification, and a four-level safety decision tree. The framework claims to outperform competitors like Anthropic, OpenAI, LangGraph, CrewAI, and AutoGen in governance and security, with a 97% automation rate and support for desktop control, multi-agent orchestration, and cross-framework adapters.", "body_md": "**M** ulti-**A** gent **R** ecursive **E** ngineering **F** ramework\n\n全球唯一以\"Agent 治理\"为核心产品定位的框架。在治理深度上碾压所有竞品（10/10 vs 0-3），将 Agent 治理作为独立的价值主张而非安全 feature。\n\nMAREF 是 Agent 世界的操作系统内核 — 管理 Agent 集群的生命周期、安全边界、状态健康和进化方向。\n\n**10 态 Gray Code 治理状态机**— 数学可证明收敛性 (6bit, 汉明距离=1)** TLA+ 形式化验证**— 5 定理证明 (Lyapunov收敛 + Sperner完备性)** CircuitBreaker**— 3连败自动锁 + HALT 吸收态 + 30s 冷却**四级安全决策树**— Rule→Mode→SafetyGate→User, 97% 自动化率** LoRA/本体双重漂移检测**— KL/JS/Hellinger 三重散度 + 人工仲裁\n\n**桌面 Agent 操控**— 截图→解析→键鼠→验证 完整闭环 (macOS/Linux/Windows)**多 Agent 任务编排**— TaskDAG 分解 + 5维 Agent 分发 + Saga 补偿事务** SubAgent 上下文隔离**— Git Worktree 式, 96% Token 节省**移动→桌面任务桥接**— mDNS 发现 + 幂等任务队列 + SSE 推送**浏览器安全操控**— Playwright + 安全域名白名单 + 认证会话管理\n\n**递归自演进引擎**— C1(观测)→C2(优化)→C3(收敛) 三循环**红蓝对抗**— 200 轮 5 阶段, 攻击强度 2.47→18.98 (7.7×)**混沌工程**— 5 类 LLM 故障注入 (延迟/错误/截断/幻觉/超时)**记忆三温框架**— Hot/Warm/Cold 三层记忆架构** Trust Engine v2**— 5 因子加权 + Goodhart 抗策略操纵检测\n\n**A2A/MCP 双协议**— A2A v0.3 + MCP 6 种传输**跨框架适配器**— AutoGen/CrewAI/LangGraph/Dify/Coze 生产级** OpenTelemetry**— Prometheus + Grafana + OTLP 全链路可观测** Serverless 运行时**— Lambda / Cloud Run 适配** TypeScript SDK**—`@maref/sdk`\n\nnpm 包\n\n```\n# 一键安装\npip install maref\n\n# 桌面操控 (dry-run 安全模式)\nmaref desktop demo\n\n# 环境诊断 (15项检查)\npython scripts/check_desktop_env.py\n\n# 治理状态查询\nmaref status\n\n# 启动 Sidecar 服务\nmaref serve --port 8000 --gui\npython\nfrom maref_lite.governance import GovernanceOverlay\nfrom maref_lite.state_machine import GovernanceState\n\noverlay = GovernanceOverlay()\noverlay._state_machine.transition(GovernanceState.OBSERVE)\noverlay._state_machine.transition(GovernanceState.ANALYZE)\nprint(overlay.get_status())\nMAREF: Agent 治理操作系统\n    ┌─────────────────────────────────────────────────────────┐\n    │  应用层 ─── LangGraph / CrewAI / AutoGen / Anthropic    │\n    │             (编排/操控/开发框架)                          │\n    │  ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─  │\n    │  治理层 ─── MAREF (本框架)                               │\n    │             · 状态机 · 熔断器 · 四级决策树               │\n    │             · 身份/信任 · 漂移检测 · 形式化验证           │\n    │  ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─  │\n    │  通信层 ─── A2A / MCP (Google/Anthropic 标准)            │\n    └─────────────────────────────────────────────────────────┘\n```\n\n| 维度 | MAREF |\nAnthropic | OpenAI | LangGraph | CrewAI | AutoGen |\n|---|---|---|---|---|---|---|\n| 治理/安全 | 10 |\n4 | 3 | 2 | 1 | 1 |\n| 形式化验证 | 10 |\n0 | 0 | 0 | 0 | 0 |\n| 漂移检测 | 9 |\n0 | 0 | 0 | 0 | 0 |\n| 桌面操控 | 8 | 9 |\n7 | 0 | 0 | 0 |\n| 编排 | 7 | 8 | 8 | 9 |\n8 | 8 |\n| 身份/信任 | 7 |\n0 | 0 | 0 | 0 | 0 |\n| 社区/生态 | 3 | 8 | 9 |\n8 | 9 |\n8 |\n\n| 类型 | 数量 | 状态 |\n|---|---|---|\n| 全量 | 4,300+ | ✅ |\n| 覆盖率 | 82% | ✅ |\n\n```\npytest tests/ -v --cov\npytest tests/desktop/ -v    # 桌面操控测试\npytest tests/chaos/ -v       # 混沌工程\n```\n\n- v0.1.0-v0.20.0: 工程基础设施 + 形式化验证 + Sidecar + 漂移 + 混沌 + A2A + Identity + 编排 + Desktop Agent → GA\n- Phase Ω (R101-R150): 50 轮自主递归演进全量补强 → v0.21.0 Final\n- v0.30.0-GA: 人机协同层 + 记忆层 + 技能市场层 + 国密 SM2/SM3/SM4-GCM + 技术白皮书\n- v1.0: 递归进化全栈 + Agent 信用评级 + 四象治理模型\n- v2.0: 元 Agent 闭包 + 碳硅共生 + 八卦治理\n\n详见 [task_plan_v0.21.0-rc_omega_50_rounds.md](/maref-org/maref/blob/main/task_plan_v0.21.0-rc_omega_50_rounds.md)\n\nApache License 2.0 — [LICENSE](/maref-org/maref/blob/main/LICENSE)", "url": "https://wpnews.pro/news/maref-formal-governance-for-multi-agent-systems", "canonical_source": "https://github.com/maref-org/maref", "published_at": "2026-06-18 10:25:49+00:00", "updated_at": "2026-06-18 10:53:31.848602+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "ai-infrastructure", "ai-research", "ai-tools"], "entities": ["Maref", "Anthropic", "OpenAI", "LangGraph", "CrewAI", "AutoGen", "Google", "Apache License 2.0"], "alternates": {"html": "https://wpnews.pro/news/maref-formal-governance-for-multi-agent-systems", "markdown": "https://wpnews.pro/news/maref-formal-governance-for-multi-agent-systems.md", "text": "https://wpnews.pro/news/maref-formal-governance-for-multi-agent-systems.txt", "jsonld": "https://wpnews.pro/news/maref-formal-governance-for-multi-agent-systems.jsonld"}}