HICOOL Developer Challenge · MTClaw Agent Track Project Plan

Project Name: BossAgents Intelligent Enterprise Brain

Project Positioning: An enterprise-level digital employee intelligent scheduling platform based on MTClaw

Participating Team: [Team Name]

Submission Date: July 2026

Table of Contents

1. Project Overview and Core Value
2. MTClaw Key Innovations
3. Technical Architecture Solution
4. Core Assets: 43 Digital Employees
5. Performance Test Data
6. Coverage of Evaluation Dimensions
7. Demonstration and Deployment Plan

1. Project Overview and Core Value

1.1 One-Sentence Positioning

"BossAgents Intelligent Enterprise Brain" — A digital employee platform running locally on MTT AIBOOK, covering all enterprise business scenarios, and achieving "fast, accurate, and decisive" scheduling based on MTClaw.

1.2 Core Value Proposition (Four-Dimensional Validation)

| Dimension | Value Description | Validation Status |

|-----------|-------------------|-------------------|

| High Availability | MTClaw enables the system to operate 100% independently even when the main service is unavailable | ✅ Validated in four scenarios |

| Fast | Routing decision 79ms, edge-side execution 80ms, P95

| Accurate | Tool recall rate 100%, routing hit coverage for 43 employees | ✅ Validated by 860 full-scale tests |

| Decisive | 43 digital employees cover 10+ domains, 75% edge-side execution ratio | ✅ Validated by 860 full-scale tests |

1.3 What MTClaw Makes Possible

The core value of MTClaw is not "calling models faster," but "enabling agents to be truly deployable in industrial scenarios."

| MTClaw Technical Feature | What We Achieved |

|--------------------------|------------------|

| High Availability Guarantee | 100% operational even when the main service is down — this is the unique capability brought by MTClaw |

| 75% On-Device Execution Ratio | 75% of tasks require no model calls, response time reduced from 500ms+ to 80ms |

| Intelligent Routing & Traffic Shaping | Complex reasoning automatically routed to the most suitable model, simple tasks handled on-device |

| Keyword Matching Routing | Tool routing completed in 16ms, no need to wait for large model inference |

II. Key Innovations of MTClaw

Core Mechanism: Finite State Machine + Function Router + Keyword Matching Acceleration

2.1 Four-Level Routing Strategy

2.2 Degradation and High Availability Mechanisms

L1关键词匹配 → L2端侧小模型 → L3端侧大模型 → L4云端大模型 → 友好错误提示
     ↓失败           ↓失败           ↓失败           ↓失败
   自动降级        自动降级        自动降级        记录日志

关键结论：当后端大模型不可用时，MTClaw仍能通过L1+L2完成全部路由决策和端侧执行，保障业务连续性。场景D（绕过MTClaw）成功率0% vs 场景A（MTClaw）成功率100%，证明高可用性归属MTClaw。

III. Technical Architecture Solution

┌─────────────────────────────────────────────────────────────────┐
│                    用户交互层（自然语言）                         │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│              MTClaw Function Router（智能分流层）                 │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐            │
│  │ 关键词匹配   │→│ 路由决策     │→│ 工具执行     │            │
│  │ (<5ms)      │  │ (79ms)      │  │ (80ms)      │            │
│  └─────────────┘  └─────────────┘  └─────────────┘            │
│       │                 │                 │                     │
│       ▼                 ▼                 ▼                     │
│  端侧小模型        端侧大模型        云端大模型                   │
│ deepseek-r1:8b    qwen3.6          SenseNova                   │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                 BossAgents 数字员工层（43 Subagents）              │
│  采购 · 供应链 · 工艺 · 数据 · 芯片 · Knowledge base · 工作流 · 分析等10+领域 │
└─────────────────────────────────────────────────────────────────┘

3.1 三层架构

┌─────────────────────────────────────────────────────────────────┐
│  平台层：SCSAi PLM底座（15年自研，100%自主源码）                    │
│  469个ItemType · 2228个Method · 2306条业务规则 · 3489条自动生成规则  │
└─────────────────────────────────────────────────────────────────┘
                               │
                               ▼
┌─────────────────────────────────────────────────────────────────┐
│  能力层：MTClaw智能调度 + 规则引擎 + 端侧模型                        │
│  路由决策79ms · 3489条自动生成规则 · 75%端侧执行比例                  │
│  六大能力：Identify · Create · Repair · Optimize · Compare · Generate                    │
└─────────────────────────────────────────────────────────────────┘
                               │
                               ▼
┌─────────────────────────────────────────────────────────────────┐
│  场景层：43个数字员工覆盖10+工业领域                                   │
│  860次全量测试100%成功 · 75%端侧执行 · Loop Engineering自循环        │
└─────────────────────────────────────────────────────────────────┘

3.2 六大能力映射

| 能力 | 说明 | 代表员工 |

|------|------|----------|

| Identify | Data identification, classification, health check | DS-SCSAI-001, DS-STAT-001, DS-KNOWLEDGE-001, DS-SCSAI-001 |

| Create | Data import, collection, document generation | DS-SCSAI-001, DS-DATA-001, DS-DOC-001, DS-CONTENT-001 |

| Fix | Data Repair, Quality Validation | DS-PROC-DATA-001, DS-DOC-001, DS-HEALTH-001 |

| Optimize | Cost Optimization, Process Optimization, Supply Chain Optimization | DS-COST-001, DS-PROCESS-OPT-001, DS-PROC-001, DS-BIZ-001 |

| Compare | Price Monitoring, Change Analysis, Valuation Comparison | DS-LOOP-001, DS-ECR-001, DS-ECR-001, DS-COST-001 |

| Generate | Content Creation, Report Generation, News Summarization | DS-CONTENT-001, DS-REPORT-001, DS-NEWS-001, DS-LOOP-001 |

IV. Core Assets: 43 Digital Employees

4.1 7 New High-Value Subagents (All Developed ✅)

4.2 Domain Coverage Panorama (36+7=43 Digital Employees)

| Domain | Employee Count | Representative Employees |

|------|--------|----------|

| IT Operations | 4 | DS-DATA-001/002, DS-OPS-001, ⟦DS-SCSAI-001⟧ |

| Procurement Management | 5 | DS-PROC-001, DS-COST-001, DS-VEN-001, DS-PROC-001 |

| Data Assets | 6 | DS-SCSAI-001, DS-DATA-001, DS-STAT-001, DS-PROC-DATA-001 |

| Engineering Change | 3 | DS-ECR-001, DS-ECR-001, DS-PROCESS-OPT-001 |

| Chip Design | 7 | DS-CHIP-001~007 |

| Business Analysis | 2 | DS-BIZ-001, DS-REPORT-001 |

| Inventory Management | 2 | DS-STOCK-001, DS-LOOP-001 |

| Content Creation | 2 | DS-CONTENT-001, DS-LOOP-001 |

| PLM Documents | 1 | DS-PLM-BRAIN-001 |

| 7 New Additions | 7 | Knowledge Base/Workflow/Data Analysis/Health/Vision/Writing/Information |

Coverage: 10+ domains fully covered, far exceeding the requirement of "at least 3 Subagents" in the challenge.

V. Performance Test Data

5.1 Full Test Overview (860 runs, 2026-07-17)

5.2 Key Performance Indicators

5.3 Four-Scenario Comparative Validation (Proving MTClaw's High Availability Attribution)

| Scenario | Scheduling Method | Availability (When Server Fails) | Conclusion |

|------|----------|----------------------|------|

| Scenario A | MTClaw + On-device Model | 100% | MTClaw runs independently |

| Scenario B | MTClaw + Cloud Model | 0% (Network-dependent) | Fails if cloud is unavailable |

| Scenario C | No MTClaw (Self-scheduling) | 0% | Depends on main service |

| Scenario D | Bypass MTClaw, Direct Connection | 0% | Proves high availability belongs to MTClaw |

Key Finding: Scenario A vs. Scenario D — Same hardware, same model, the only variable is "whether MTClaw is used." Result: 100% vs. 0%, proving high availability is the unique value brought by MTClaw.

5.4 End-Side Execution Ratio Optimization Trajectory

| Stage | End-Side Execution Ratio | Average Response | Description |

|------|------------|----------|------|

| Before Optimization | 5.6% (1/16) | 500ms+ | 15 tool calls to large model |

| After Optimization | 75% (12/16) | 80ms | Deterministic calculations changed to local scripts |

| Change | +69.4% | -84% | Cost is almost zero |

Conclusion: It's not about tuning models faster, but about making 75% of tasks require no model tuning at all.

VI. Review Dimension Coverage Description

| Scoring Dimension | Official Requirement | Our Achievement | Evidence |

|----------|----------|------------|------|

| Fast | Response Time/Speedup Ratio | Routing 79ms/Edge 80ms, P95

| Accurate | Recall Rate/Accuracy/Domain Expertise | 100% recall rate, all 43 employees matched | 860 full test reports |

| Robust | Scenario Breadth/Number of Subagents/Productization | 43 employees/10+ domains/75% edge-side execution | Employee list + test reports |

| Ecosystem Value | Scenario Value/Business Model/Adaptation | Enterprise-level scenarios + standard APIs + MTClaw collaboration | Solution docs + deployment guide |

VII. Demo and Deployment Plan

7.1 Deployment Environment

7.2 Deployment Steps (Ubuntu/AIBOOK)

# 1. 安装Ollama
curl -fsSL https://ollama.com/install.sh | sh && ollama pull deepseek-r1:8b

# 2. 安装项目
cd bossagents && npm install && pip install -r requirements.txt

# 3. 启动服务
ollama serve & cd MTClaw && python -m function_router.server & npm start

7.3 交付物清单

| 交付物 | 内容 | 状态 |

|--------|------|------|

| Planning Document | This PDF document | ✅ Completed |

| Demo Video | 5-stage, 7-minute continuous demo | To be recorded |

| Project Code | Full source code + installation script | ✅ Completed |

VIII. Ecosystem Adaptation and Commercial Value

8.1 Ecosystem Adaptation Value

8.2 Business Value

IX. Risks and Countermeasures

Document Version: v2.0 (Optimized Edition)

Creation Date: July 17, 2026

Update Date: July 18, 2026

Applicable Competition: HICOOL Developer Challenge · Agent Track