BossAgents · GOAI New Intelligent Base Track Technical Architecture Document

Mapping Relationship with AgentTeams

The BossAgents position orchestration system naturally aligns with the AgentTeams layered architecture:

| AgentTeams Level | BossAgents Implementation | Responsibility |

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

| Manager | lite-scheduler.js + AI Workbench | Task reception, global scheduling, result aggregation |

| Team Leader | 3 Positions (POS-GOAI/POS-PROCURE/POS-INSPECT) | Scenario-based task orchestration, pipeline management |

| Worker | 10 GOAI Agents (Independent Prompt + Independent Toolset + Independent Context) | Executing specific subtasks |

Currently, all 3 positions are based on the AgentTeams sequential orchestration mode (buildChain() in position-runtime.js),

supporting hierarchical scheduling: Manager-level task distribution → Team Leader-level pipeline orchestration → Worker-level step-by-step execution.

System Architecture

┌─────────────────────────────────────────────────────────┐
│                    L5 开放层                              │
│  API Key鉴权(ba_前缀) · OpenAPI 3.0规范 · 在线调试文档    │
│  27个API端点 · 71员工逐个可调用 · 双轨制鉴权              │
├─────────────────────────────────────────────────────────┤
│                    L4 Interaction Layer                              │
│  HITS工作台 · AI对话 · 小程序 · 飞书 · 人在回路审核       │
│  15个岗位均配人类负责人(human字段)                        │
├─────────────────────────────────────────────────────────┤
│                    L3 编排层                              │
│  ┌──────────────┐ ┌──────────────┐ ┌──────────────┐     │
│  │ POS-GOAI     │ │ POS-PROCURE  │ │ POS-INSPECT  │     │
│  │ 3Agent串行   │ │ 4Agent串行   │ │ 3Agent串行   │     │
│  └──────────────┘ └──────────────┘ └──────────────┘     │
│  + 12个业务岗位(parallel) + 4个跨岗位编队(TaskForce)     │
│  position-runtime.js · CollaborationChain               │
│  事件总线(position-events.js) · SSE实时流 · 触发器热更新  │
├─────────────────────────────────────────────────────────┤
│                    L2 能力层                              │
│  ┌─────────────────────────────────────────────────┐    │
│  │ 6大原子能力: identify/validate/repair/          │    │
│  │              optimize/compare/generate/         │    │
│  │              create/import/inspect/valuate      │    │
│  ├─────────────────────────────────────────────────┤    │
│  │ 71个数字员工 + 2个LLM协同Worker Skill             │    │
│  │ goai-agent-worker · procurement-chain-worker    │    │
│  ├─────────────────────────────────────────────────┤    │
│  │ Agent Card接入 · trace_span协议                  │    │
│  │ dispatch/callback/audit 三接口                    │    │
│  └─────────────────────────────────────────────────┘    │
├─────────────────────────────────────────────────────────┤
│                    L1 底座层                              │
│  ┌────────────┐ ┌────────────┐ ┌────────────┐          │
│  │ 关系引擎   │ │ 规则引擎   │ │ SCSAI PLM  │          │
│  │ 73.9万关系 │ │ 2306条规则 │ │ 13000对象  │          │
│  └────────────┘ └────────────┘ └────────────┘          │
│  ┌────────────┐ ┌────────────┐                          │
│  │ MTCLAW     │ │ SmartLLM   │                          │
│  │ 昇腾NPU    │ │ Router     │                          │
│  │ 本地推理   │ │ 降级链     │                          │
│  └────────────┘ └────────────┘                          │
│  audit-logger · sqlite-compat · open-platform           │
└─────────────────────────────────────────────────────────┘

行业×岗位×员工三层正交坐标系

BossAgents 的核心架构创新：行业、岗位、员工三层完全解耦。

| Dimension | Role | Quantity | Variability | Example |

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

| Industry | Knowledge Coordinate System | 5 | Fixed Set | Phosphorus Chemical/Liquor/Machinery/General Machinery/Electronics Manufacturing |

| Position | Authority & Responsibility Entity | 15 | Adjustable | POS-OPERATOR(Operator Assistant)/POS-QA(Quality Administrator Assistant) |

| Employee | Capability Atom | 71 | Universal | DS-PROC-001(Procurement Assistant)/DS-PROC-DATA-001(Data Repair Specialist) |

Decoupling Benefits: Cost is N+M+K instead of N×M×K. The same group of 71 employees switches contexts based on the industry knowledge base and changes collaboration modes based on role orchestration. Adding a new industry only requires injecting an industry knowledge base, with no need to hire new employees; adding a new role only requires orchestrating existing employees, with no need to develop new Agents.

Industry Consumption Mechanism: 4 employees actually consume data along the industry dimension (DS-KBPIPE-001 Industry Knowledge Base Pipe Operator, DS-DOC-001 Document Digital Employee, DS-KNOWLEDGE-001 Knowledge Base Assistant, DS-PROC-DATA-001 Process Data Repair Operator), while the remaining employees are triggered based on role orchestration.

Human and Digital Employee Collaboration Mechanism

1. Human-in-Charge for Each Role (Human Layer)

Each role configures a human field to specify the human responsible:

| Role | Human-in-Charge | Employee Count |

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

| POS-OPERATOR Operator Assistant | Operator / Team Leader | 4 |

| POS-QA Quality Administrator Assistant | Quality Administrator / Quality Manager | 5 |

| POS-PROCURE Procurement Specialist Assistant | Procurement Specialist / Procurement Manager | 5 |

| POS-BOSS Boss Assistant | Business Owner / General Manager | 9 |

| ... | ... | ... |

2. Human-in-the-Loop

• Formation Manual Checkpoints: Cross-role formations support requireHumanApproval: true. Execution automatically suspends at manual sign-off nodes, resuming only after human approval via POST /api/digital-staff/taskforce/resume.
• High-Risk Operation Mandatory Confirmation: Operations like procurement exceeding threshold amounts, ECR approval, Part release, etc., enforce human-in-the-loop.
• Real-time event push: Event bus SSE stream (GET /api/digital-staff/event/stream) pushes events in real time, allowing for human intervention at any time.

3. Cross-Role Task Force

Four task forces adopt DAG layering + shared blackboard + dual-layer conflict detection:

| Task Force | Trigger Scenario | DAG Structure |

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

| TF-COMPLAINT | Customer complaint | Quality → Production → Procurement → Customer Service |

| TF-LINE-STOP | Production line stoppage | Equipment → Production → Procurement → Scheduling |

| TF-NEW-PRODUCT | New product introduction | Design → Process → Quality → Procurement |

| TF-COST-ALERT | Cost warning | Finance → Procurement → Production → Quality |

Conflict detection: Hard collisions (mutual exclusion of the same resource) + Semantic disagreements (LLM judges contradictory suggestions) → Automatic arbitration or escalation to human.

Open Platform

Dual-track Authentication

API Endpoints (27)

Online Debugging Documentation

GET /api/open/docs returns a complete Swagger UI-style documentation page, where 71 digital employees can be invoked one by one for testing, and each of the 27 API endpoints includes a "Try it" button.

Core Modules

1. Position Orchestration Engine (position-runtime.js)

• Function: Industry→Position→Employee three-layer model, supporting serial/parallel orchestration
• Serial Mode: Upstream results are automatically passed to downstream Agents via _upstreamResult
• Collaboration Chain: buildChain() generates visual collaboration chain data
• Unified Result Contract: normalizeResult() standardizes output format for all Agents

2. LLM Collaboration Worker

| Worker | File | Purpose |

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

| goai-agent-worker | workers/goai-agent-worker.js | GOAI scenario + inspection loop, 6 independent systemPrompts |

| procurement-chain-worker | workers/procurement-chain-worker.js | Procurement loop, 4 independent systemPrompts |

Design principles:

• Each Agent has an independent systemPrompt, with completely distinct roles.
• Each Agent has independent context, only viewing structured data passed from upstream.
• Upstream results are passed as structured JSON, not natural language relay.
• LLM returns are forced into JSON format, parsed with extractJson() for error tolerance.

3. Skill Engineering Encapsulation

BossAgents encapsulate key capabilities into two types of reusable Skills:

1. Worker Skill (Internal Skill)

| Skill Name | Worker | Encapsulated Process | Reusable Scenarios |

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

| Procurement Closed-loop Skill | procurement-chain-worker | Inquiry→Comparison→Evaluation→Approval | Any role requiring procurement approval |

| GOAI Collaboration Skill | goai-agent-worker | Task decomposition→Procurement collaboration→Equipment collaboration | Any composite instruction scenario |

| Inspection Closed-loop Skill | goai-agent-worker | Inspection→Repair→Re-inspection | Any data quality governance scenario |

2. Agent Card Skill (External Integration Skill)

• Capability tags (capabilities) are declared via Agent Card JSON to enable skill discovery and matching.
• Any third-party Agent can be integrated as a Skill by implementing the dispatch/callback interface.
• Interface protocol: trace_span (refer to Agent Card specification document).
• Skills are reusable across scenarios: The Procurement Closed-loop Skill can be referenced in any role requiring procurement.

3. Audit Engine (audit-logger.js)

• trace_id: A global identifier that persists throughout a collaborative closed-loop process.
• span_index: Identifies the sequence number of each step in the closed loop.
• Fields: timestamp, trace_id, span_index, agent_id, action, detail, metadata
• Query: GET /api/agents/audit/{trace_id} returns the complete span chain.
• AgentLoop Integration: The structured trace_id/span data output by the audit-logger can seamlessly integrate with the AgentLoop observability platform, supporting full-link Trace visualization, Agent performance evaluation, and self-evolution tuning.

4. Agent Interface (goai-demo.js)

5. Rule Engine (rule-engine.js)

• 2306 deterministic rules, zero hallucination, zero Token consumption.
• Supports 5 check types: model/instance/relationship/alignment/custom.
• Automatic repair (auto) + Manual repair (manual) grading.
• Collaboration with LLM: Rule engine takes priority, LLM serves as fallback.
• High-risk operations (e.g., purchase approval amount exceeding threshold) require manual confirmation (human-in-the-loop), supporting operation rollback. All rollback operations are also recorded in the audit chain.

6. SmartLLMRouter + MTCLAW

• Unified LLM scheduling entry point, supporting multi-model fallback chain
• Current model: deepseek-v4-flash (sensenova endpoint)
• Timeout 120s, supports transparent transmission of response_format and max_tokens
• MTCLAW scheduling engine: When MTCLAW_ENABLED=true, SmartLLMRouter prioritizes MTCLAW local inference (Ascend NPU accelerated, port 18790), falling back to remote API
• Edge-side privatization: MTCLAW local inference means LLM stays within the enterprise network, meeting manufacturing data security and compliance requirements
• Fallback chain: MTCLAW local → sensenova remote → backup model, automatic seamless switching

Data Scale

Tech Stack

| Layer | Technology |

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

| Frontend | Vue 3 + Vite + Pinia |

| Backend | Node.js (Native HTTP) |

| Database | SQLite (better-sqlite3) + SCSAI AML |

| AI | DeepSeek v4 Flash (sensenova) |

| Rule Engine | Self-developed (rule-engine.js) |

| Audit | Self-developed (audit-logger.js + trace_id/span) |

| RAG | SCSAI PLM AML Knowledge Base (13,000+ object models) as RAG knowledge source, automatically retrieved and enhanced during Agent invocation |

| MCP | The system supports connecting external tools via the MCP (Model Context Protocol). Currently, similar tool discovery and invocation capabilities are implemented through Agent Card capabilities.

Open Source Plan

• Open Source License: Apache License 2.0
• Open Source Scope:
• ✅ Agent Card Specification (JSON Schema)
• ✅ trace_span Communication Protocol (dispatch/callback/audit three interfaces)
• ✅ position-runtime.js Position Orchestration Engine
• ✅ audit-logger.js Audit Engine
• Not Open Source: Industrial knowledge graph data (739,000 relationships), rule engine rule base (2,306 rules), business configuration data
• Experience Entry: ylxt.chat offers login-free access to all 71 digital employees
• Long-term Plan: Continuously open more Worker Skill interfaces, supporting community contributions of custom Skills

Deployment

# 启动
node server.js

# 验证
curl https://ylxt.chat/api/health

# 演示
curl https://ylxt.chat/api/goai-demo/positions
curl -X POST https://ylxt.chat/api/goai-demo/collaboration/goai \
  -H "Content-Type: application/json" \
  -d '{"instruction":"3号泵备件库存不足"}'