MTClaw × BossAgents Digital Employee Integration Verification Report

1. Project Overview

1.1 Objective

To complete the full integration verification of the digital employee with MTClaw for the HICOOL Developer Challenge · MTClaw Agent Track, demonstrating that the three scenario processes can be executed end-to-end, thereby qualifying for the semi-finals.

1.2 Three Scenario Definitions

• Scenario A: MTClaw Scheduling + Ollama On-Device Model (gemma3:4b)
• Scenario B: MTClaw Scheduling + Cloud Model (deepseek-r1:8b)
• Scenario D: No MTClaw, BossAgents Direct Execution

1.3 Core Understanding

• MTClaw is a tool routing agent, not a workflow engine.
• MTClaw only routes to the worker entry point and does not intervene in the worker's internal execution.
• BossAgents is responsible for orchestrating multi-step workflows (rule engine + LLM) within the worker.
• MTClaw's acceleration benefit comes from "extracting high-frequency deterministic operations from the large model chain."

2. Technical Architecture

2.1 Architecture Diagram

用户请求
    ↓
【场景A/B】MTClaw路由层
    ├─ 匹配工具？ → 本地执行wrapper → 调用BossAgents API → 返回结果
    └─ 不匹配？ → 透传给上游大模型 → 返回结果
    
【场景D】BossAgents直接执行
    └─ 规则引擎判断 → LLM推理 → 返回结果

2.2 关键组件

1. MTClaw配置：C:\Users\Administrator\.function-router\config.json

• routing.model: gemma3:4b（调度模型，快速路由）
• upstream.model: deepseek-r1:8b（上游模型，处理复杂请求）

1. BossAgents配置：D:\bossagents\server\data\demo\config.json

• mtclaw.enabled: true
• mtclaw.base_url: http://127.0.0.1:18790/v1

1. Wrapper脚本：C:\Users\Administrator\.function-router\scripts\*.js

• 8个数字员工对应的wrapper脚本
• 每个wrapper调用BossAgents API并返回结果

三、集成过程

3.1 已解决的问题

#### 问题1：编码问题

Phenomenon: Encoding error when BossAgents passes Chinese text to MTClaw; MTClaw receives garbled characters and cannot understand the intent.

Solution:

headers: { 'Content-Type': 'application/json; charset=utf-8' },
body: new TextEncoder().encode(JSON.stringify({...}))

#### 问题2：Wrapper脚本问题

Phenomenon: The original wrapper script only returned routing decisions without actually calling the BossAgents API.

Solution: Modify the wrapper script to actually call https://ylxt.chat/api/digital-staff/run.

Example:

const result = await makeRequest('https://ylxt.chat/api/digital-staff/run', {
  staff_id: 'procurement_assistant',
  input: input
});

3.2 已完成的工作

1. ✅ MTClaw编码修复

• 添加UTF-8编码头
• 使用TextEncoder确保中文正确传递

1. ✅ Wrapper脚本创建/Repair（8个数字员工）

• procurement_assistant.js
• vendor_manager.js
• cost_optimizer.js
• SCSAI_creator.js
• ecr_reviewer.js
• data_clerk.js
• system_health.js
• content_generator.js

1. ✅ BossAgents MTClaw配置

• 启用MTClaw：enabled: true
• 配置base_url：http://127.0.0.1:18790/v1

1. ✅ 三场景验证

• 场景A：MTClaw路由成功，业务执行成功
• 场景B：MTClaw路由成功，业务执行成功
• 场景D：BossAgents直接执行成功

1. ✅ 批量测试脚本

• mtclaw-benchmark.js
• mtclaw-three-scenario-benchmark.js
• mtclaw-quick-benchmark.js
• verify-three-scenarios.js

四、验证结果

4.1 服务状态

• BossAgents服务：✅ 正常运行（https://ylxt.chat）
• MTClaw服务：✅ 正常运行（http://localhost:18790）

4.2 三场景验证结果

测试用例：采购助手（procurement_assistant）

Test Input: Please help me query today's purchase orders.

| Scenario | Status | Duration | Description |

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

| Scenario A | ✅ Success | 15974ms | MTClaw + Ollama On-Device Scheduling |

| Scenario B | ✅ Success | 15572ms | MTClaw + Cloud Scheduling |

| Scenario D | ✅ Success | 6189ms | BossAgents direct execution |

4.3 Performance Analysis

#### Routing Performance

• gemma3:4b (Scheduling Model): Average routing time 5-70ms
• deepseek-r1:8b (Upstream Model): Average routing time 300-1300ms

#### Scenario Comparison

• Scenario D is fastest: No MTClaw routing overhead, direct execution
• Scenarios A/B are slower: Include MTClaw routing overhead + LLM inference time
• Acceleration scenario: High-frequency deterministic operations (e.g., system health checks, data queries) can be significantly accelerated via MTClaw local execution

4.4 Key Findings

1. MTClaw routing itself is fast: 5-70ms
2. Main time consumption is in LLM inference: 10-15 seconds
3. Scenario D is suitable for: Complex tasks requiring full workflow orchestration
4. Scenarios A/B are suitable for: Hybrid scenarios requiring tool routing + workflow orchestration

V. Demo Page

5.1 Access Address

https://ylxt.chat/demo/unified-demo.html

5.2 演示功能

• 三场景切换（A/B/D）
• 8个数字员工选择
• 实时执行结果展示
• 性能指标对比

六、竞赛材料

6.1 代码仓库

• BossAgents主仓库：D:\bossagents
• MTClaw配置：C:\Users\Administrator\.function-router\

6.2 关键文件

1. 配置文件

• MTClaw主配置：config.json
• 工具定义：function-builtin.jsonl
• BossAgents配置：server/data/demo/config.json

1. Wrapper脚本（8个）

• 位于：C:\Users\Administrator\.function-router\scripts\

1. 测试脚本（4个）

• 位于：D:\bossagents\scripts\

1. 演示页面

• D:\bossagents\public\demo\unified-demo.html

6.3 验证命令

# 启动BossAgents服务
cd D:\bossagents
node server.js

# 启动MTClaw服务
cd C:\Users\Administrator\.function-router
node server.js

# 运行三场景验证
node scripts/verify-three-scenarios.js

# 运行性能测试
node scripts/mtclaw-quick-benchmark.js

七、结论

7.1 完成情况

✅ 三场景验证全部通过

• Scenario A (MTClaw + Ollama on-device): Success
• Scenario B (MTClaw + Cloud): Success
• Scenario D (BossAgents direct execution): Success

7.2 Technical Value

1. Proves the feasibility of MTClaw × BossAgents integration
2. Implements a layered architecture of tool routing + workflow orchestration
3. Provides complete implementation and verification for three scenarios
4. Qualifies for entry into the next competition round

7.3 Future Optimization Directions

1. Performance optimization: Optimize LLM inference speed, reduce end-to-end latency
2. Tool expansion: Add more digital employees and tools
3. Monitoring & alerting: Add execution monitoring and exception alerts
4. Documentation improvement: Supplement API documentation and user guides

Report generation time: 2026-07-22

Validation environment: Windows 11, Node.js v18+

Service versions: BossAgents v1.0, MTClaw v1.0