BossAgentsBossAgents · GOAI New Intelligence Foundation Track Preliminary Submission
Project Positioning
BossAgents — Digital Workforce Orchestration Open Platform: Enables multiple AI agents to autonomously complete the full-chain closed loop of spare parts procurement and equipment maintenance in manufacturing workshops through deterministic rules + LLM collaboration. Any enterprise can directly invoke the capabilities of 71 digital employees via API, and any agent can join the collaborative network via Agent Card.
Core Narrative: Trustworthy AI is not a slogan; it's an architecture. We do not promise AI is always correct, but we guarantee every decision is traceable, auditable, and correctable. BossAgents is not just a product, but an open platform — open interfaces, open protocols, open ecosystem.
Three Demonstration Scenarios
Scenario 1: Workshop Composite Instruction Collaboration (POS-GOAI)
Script: "Insufficient spare parts inventory for Pump No. 3, currently only 2 pieces remain with a safety stock of 10 pieces. Also, the maintenance for Centrifugal Pump No. 3 is about to expire and requires scheduled maintenance."
| Agent | ID | Role | Independent Prompt | Independent Toolset | Independent Context |
|-------|-----|------|-----------|-----------|-----------|
| Workshop Supervisor Assistant | DS-GOAI-SCHED | Task Decomposition | ✅ Understand composite instruction → Purchase request + Maintenance reminder | LLM invocation, task decomposition, priority judgment | User's original instruction |
| Procurement Assistant | DS-GOAI-PROC | Procurement Collaboration | ✅ Generate supplier comparison + Purchase order | LLM invocation, supplier comparison, purchase order generation | Upstream purchase request |
| Equipment Manager Assistant | DS-GOAI-EQUIP | Equipment Collaboration | ✅ Generate maintenance plan + Work order | LLM invocation, maintenance plan generation, work order dispatch | Upstream maintenance reminder |
Collaboration Mode: Serial (sequence), upstream results automatically passed via _upstreamResult
Measured Data (2026-08-11):
- Total Duration: 18.4s
- Agent1(Workshop Supervisor): 6.6s → Purchase request(Seal SEAL-P003-08) + Maintenance reminder(Regular maintenance for Centrifugal Pump No.3)
- Agent2 (Procurement): 4.8s → Compare quotes from 3 suppliers → Select Company A (1-day lead time / Excellent rating) → Purchase order
- Agent3 (Equipment): 6.0s → Maintenance plan (5 task steps) + Repair work order (Assign to Master Zhang)
API calls:
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，当前仅剩2件安全库存10件，同时3号离心泵保养即将到期需要安排定期保养"}'
场景二：采购闭环（POS-PROCURE-CHAIN）
话术："采购伺服电机SM-750A 50台，预算90万，15天内到货"
| Agent | ID | Role | Independent Prompt |
|-------|-----|------|-----------|
| Inquiry Assistant | DS-PROC-CHAIN-001 | Inquiry & Preliminary Comparison | ✅ Generate quotes from 3 suppliers |
| Cost Analysis | DS-COST-CHAIN-001 | Cost Baseline Comparison | ✅ Historical baseline comparison + Comprehensive score |
| Supplier Evaluation | DS-VEN-CHAIN-001 | Performance Verification | ✅ Performance grading + Risk assessment |
| ECR Review | DS-ECR-CHAIN-001 | Compliance Check | ✅ Budget/Supplier/Process compliance + Approval conclusion |
Collaboration mode: Sequential (sequence), 4-step closed loop: Inquiry → Comparison → Evaluation → Approval
Measured data (2026-08-11):
- Total duration: 22.1s
- Agent1 (Inquiry): 4.2s → 3 quotes (8200/8800/9100 RMB)
- Agent2 (Cost): 7.8s → Historical baseline comparison + Comprehensive scores (92/85/78)
- Agent3 (Evaluation): 5.8s → Performance verification (Grade A / On-time rate 98%) + Low risk
- Agent4 (Review): 3.9s → Compliance passed + Approved Supplier A + Amount 410,000 RMB
API calls:
curl -X POST https://ylxt.chat/api/goai-demo/collaboration/procure \
-H "Content-Type: application/json" \
-d '{"instruction":"采购伺服电机SM-750A 50台，预算90万，15天内到货"}'
场景三：巡检修复闭环（POS-INSPECT-LOOP）
话术："检查Part数据质量，发现并修复问题"
| Agent | ID | Role | Independent Prompt |
|-------|-----|------|-----------|
| Data Inspector | DS-INSPECT-CHAIN-001 | Five-Dimensional Scoring | ✅ Completeness/Accuracy/Timeliness/Availability/Consistency |
| Data Repairer | DS-REPAIR-CHAIN-001 | Automatic Repair | ✅ Generate repair plan based on issue list |
| Re-inspector | DS-REINSPECT-CHAIN-001 | Re-inspection Confirmation | ✅ Verify fix effectiveness + closed-loop status determination |
Collaboration mode: Sequential (sequence), 3-step closed loop: Inspection → Fix → Re-inspection
Measured data (2026-08-11):
- Total duration: 12.5s
- Agent1 (Inspection): 5.4s → Five-dimensional score (Completeness 72/Accuracy 85/Timeliness 60/Usability 90/Consistency 68) + 12 issues
- Agent2 (Fix): 2.8s → 8 auto-fixes + 4 manual → Expected improvement: 18 points
- Agent3 (Re-inspection): 4.1s → Post-fix score 90 (improved 18 points) → loop_status=closed → Closed-loop confirmation
API calls:
curl -X POST https://ylxt.chat/api/goai-demo/collaboration/inspect \
-H "Content-Type: application/json" \
-d '{"instruction":"检查Part数据质量，发现并修复问题"}'
评审标准覆盖
维度一：场景价值与可复制性（25%）
| 评分点 | 覆盖方式 |
|--------|---------|
| Real-world scenario | Manufacturing enterprise workshop spare parts procurement + equipment maintenance, non-fictional demo |
| Replicability | 3 GOAI roles support 5 industries (phosphorus chemical/liquor/machinery/general machinery/electronics manufacturing), 12 business roles cover the full manufacturing chain |
| Cross-industry data | SCSAI PLM foundation contains 13,000+ object models, 739k native relationships |
| Scenario Depth | From single query to 4-step approval closed loop, from inspection to repair and re-inspection closed loop |
| Industry × Role × Employee Three-Layer Orthogonal Coordinate System | Industry = Knowledge coordinate system (fixed set of 5 industries) | Role = Authority and responsibility subject (15, adjustable) | Employee = Capability atom (71, universal). After decoupling, cost is N+M+K instead of N×M×K. The same group of employees can switch contexts based on industry knowledge bases and switch collaboration modes based on role orchestration |
Dimension 2: Multi-Agent Collaboration and Autonomous Closed Loop (25%)
| Scoring Point | Coverage Method |
|---|---|
| ≥3 Agents with different functions | 10 Agents (3+4+3), belonging to 3 different roles, with completely distinct functions |
| Independent Prompt | Each Agent has an independent systemPrompt, with no role overlap |
| Independent Toolset | 2 Worker Skills: procurement-chain-worker (4 roles for procurement closed loop) vs goai-agent-worker (3 roles for GOAI collaboration + 3 roles for inspection closed loop). Different roles within the same Worker are distinguished by independent systemPrompts |
| Independent Context | In serial mode, each Agent only sees the structured data passed from upstream, without sharing the original conversation. |
| Autonomous Closed Loop | Inspection → Repair → Re-inspection (loop_status=closed), Purchase → Approval (approval_decision=Approved) |
| Message Passing | Upstream results are passed via _upstreamResult structured data, not through natural language relay. |
| Human-Digital Employee Collaboration | All 15 roles are assigned a human supervisor (human field). Humans participate as team members in key approvals. Four cross-role taskforces support requireHumanApproval manual checkpoints—execution automatically pauses at human sign-off nodes, resuming after human POST /taskforce/resume. High-risk operations (purchases exceeding thresholds, ECR approvals) enforce human-in-the-loop. |
| Cross-Role Taskforces | Four taskforces (TF-COMPLAINT/TF-LINE-STOP/TF-NEW-PRODUCT/TF-COST-ALERT) use DAG layering + shared blackboard + hard collision/semantic divergence dual-layer conflict detection. Taskforce suspension state is persisted; resumes after human-in-the-loop sign-off. |
Dimension 3: Skill Engineering System (25%)
| Scoring Point | Coverage Method |
|--------|---------|
| Skill Engineering Encapsulation | 2 Worker Skills encapsulated as reusable Skills: procurement-chain-worker = Purchase Closed-Loop Skill (inquiry → comparison → evaluation → approval, 4 independent systemPrompts); goai-agent-worker = General Collaboration Skill (GOAI 3 roles + Inspection Closed-Loop 3 roles, 6 independent systemPrompts) |
| Skill Reusability Across Scenarios | The goai-agent-worker serves both the GOAI collaboration and inspection closed-loop roles simultaneously; the procurement-chain-worker can be referenced in any position requiring procurement. |
| Skill Interface Standard | Three interfaces: dispatch + callback + audit, following the trace_span protocol. |
| Ecosystem Reuse | Agent Card declarative access; third-party Agents are plug-and-play. |
| Agent Card Specification | Minimal JSON declaration of capabilities and interfaces (see below). |
| Deterministic Rules | 2306 rule engine deterministic execution, zero hallucination, zero Token. |
| RAG Capability | SCSAI PLM AML knowledge base (13,000+ object models) as RAG knowledge source, automatically retrieved and enhanced when Agent is invoked. |
| MCP Support | The system supports integrating external tools via the MCP protocol; currently, similar capability is achieved through Agent Card capabilities. |
| MTCLAW Scheduling Engine | Ascend NPU-accelerated local LLM scheduling engine (MTCLAW_ENABLED=true). SmartLLMRouter prioritizes MTCLAW local inference, with fallback to remote APIs. For edge-side private deployment, data stays within the enterprise, and LLM inference is completed on local NPUs, meeting manufacturing data security requirements. |
| Event-Driven Skill | Event bus (position-events.js) supports trigger hot updates + SSE real-time fan-out. Automatically triggers procurement Skill when inventory falls below the safety line, without manual initiation. |
Agent Card Specification:
{
"agent_id": "bossagents-procurement-001",
"name": "BossAgents采购闭环",
"capabilities": ["inquiry", "cost_analysis", "vendor_evaluation", "compliance_check"],
"endpoint": "https://ylxt.chat/api/agents/dispatch",
"auth": "api_key",
"protocol": "trace_span"
}
Skill接口：
| Interface | Method | Description |
|------|------|------|
| Task Dispatch | POST /api/agents/dispatch | Carries customer_id + trace_id |
| Result Callback | POST /api/agents/callback | Return result + span data |
| Audit Query | GET /api/agents/audit/{trace_id} | Query complete audit trace by trace_id |
Dimension 4: Engineering Implementation and Security Audit (20%)
| Scoring Point | Coverage Method |
|--------|---------|
| Full-Link Audit | trace_id spans start→N agents→complete, each span records input/output/decision/timestamp |
| Observability | GET /api/agents/audit/{trace_id} real-time query of complete span trace for any trace |
| AgentLoop Integration | Audit data can be integrated with the AgentLoop observation platform, supporting performance evaluation and Agent self-evolution tuning |
| Data Isolation | customer_id isolation, each customer's Agent can only access authorized scope |
| Engineering Implementation | SCSAI PLM 15-year industrial foundation, 13000+ object models already running in production environment |
| Deployment Method | Edge-side private deployment, data stays within the enterprise |
| Rollback Mechanism | High-risk operations require manual confirmation (human-in-the-loop), support operation rollback, rollback operations recorded in audit trace |
| Open Platform | 27 API endpoints + API Key authentication (ba_ prefix) + rate limiting (free/pro/enterprise tiers) + call audit. OpenAPI 3.0 specification (/api/open/spec) + online debugging documentation (/api/open/docs, 71 employees can call individually). Dual-track system: with Key for authentication+rate limiting+audit, without Key maintains internal logic unchanged |
Audit Trace Example:
trace_id: "trace-procure-1786449652571"
├── span[0]: POS-PROCURE-CHAIN 闭环启动
│ └── detail: "指令: 采购伺服电机SM-750A 50台"
├── span[1]: DS-PROC-CHAIN-001 询价完成
│ └── llm_ms: 4193, data_keys: inquiry,quotations
├── span[2]: DS-COST-CHAIN-001 成本分析完成
│ └── llm_ms: 7794, data_keys: cost_analysis,scoring
├── span[3]: DS-VEN-CHAIN-001 供应商评估完成
│ └── llm_ms: 5790, data_keys: vendor_evaluation,risk_assessment
├── span[4]: DS-ECR-CHAIN-001 ECR审核完成
│ └── llm_ms: 3877, data_keys: compliance_check,approval_decision
└── span[5]: POS-PROCURE-CHAIN 闭环完成
└── duration_ms: 22069
维度五：开放与开源贡献（5%）
| 评分点 | 覆盖方式 |
|--------|---------|
| Open Source License | Apache License 2.0 |
| Open Source Content | Agent Card specification, trace_span protocol, job orchestration engine, audit engine |
| Open Platform | API Key authentication + OpenAPI 3.0 specification + online debugging documentation (/api/open/docs, 71 employees can try each API) + 27 API endpoints |
| Experience Entry | ylxt.chat provides login-free access to all 71 digital employees |
| Long-term Plan | Continuously open more Worker Skill interfaces, supporting community contributions of custom Skills |
Technical Architecture (Based on AgentTeams)
BossAgents adopts the AgentTeams layered architecture for multi-agent collaboration:
| AgentTeams Layer | BossAgents Implementation | Responsibilities |
|----------------|----------------|------|
| Manager | lite-scheduler.js + AI Workbench | Receives tasks, global scheduling, result aggregation |
| Team Leader | 15 positions (3 GOAI demos + 12 business positions) | Contextual task orchestration, link management |
| Worker | 71 digital employees (independent Prompt + independent toolset + independent context) | Executes specific subtasks |
L5 开放层：API Key鉴权 + OpenAPI 3.0 + 在线文档（71员工逐个可调用）
L4 Interaction Layer：HITS工作台，人作为团队成员参与审核（人在回路）
L3 编排层：15个岗位 + 4个跨岗位编队（DAG+共享黑板+人工卡点）+ 事件总线（SSE实时流）
L2 能力层：6大原子能力 × 71个数字员工 × 2个LLM协同Worker Skill × 事件触发器
L1 底座层：关系引擎（73.9万关系图谱）+ 规则引擎（2306条确定性规则）+ MTCLAW（昇腾NPU本地推理）+ SCSAI PLM
行业×岗位×员工三层正交坐标系：
- Industry = Knowledge coordinate system (5 industries: Phosphorus Chemical/Alcohol/Machinery/General Machinery/Electronics Manufacturing), the same group of employees switch context based on industry knowledge base
- Position = Authority and responsibility entity (15 positions, each assigned a human responsible person), switching collaboration mode based on position orchestration
- Employee = Capability atom (71, general-purpose), decoupled cost is N+M+K instead of N×M×K
Human-Digital Employee Collaboration Mechanism:
- Each position configures a human field (e.g., "Workshop Director/Team Leader"), humans act as team members, not observers
- Cross-functional team support requires human approval: true, automatically suspends at manual sign-off nodes
- High-risk operations (procurement exceeding thresholds, ECR approval, Part release) enforce human-in-the-loop
- Event bus real-time push (SSE), humans can intervene at any time
Core philosophy: Relationships as agreements · Agents as digital employees · Humans as team members
Data scale (time barrier that cannot be rushed)
| Dimension | BossAgents | General LLM Solutions | Traditional PLM |
|---|---|---|---|
| Object Model | 13,000+ | Needs to be built from scratch | Thousands |
| Relationship Graph | 739,000 | No relationship awareness | BOM hierarchy only |
| Rule Engine | 2,306 rules | Pure LLM hallucination | Hard-coded |
| Digital Employees | 71 (including 10 GOAI Agents) | No deterministic execution | No AI capability |
| Roles | 15 + 4 cross-functional teams | No orchestration | No collaboration |
| Industry | Validated in 5 industries | No industry adaptation | Single-industry customization |
| Audit Trail | Full-link trace_id/span | Black box | Dispersed logs |
| Deployment | Edge-side privatization | Cloud-only | On-premises |
Team
- CEO Fu Bo: 27 years of serial entrepreneurship, mentored by "Father of Modern Marketing" Kotler
- CTO Fang Yunchao: Beihang University Computer Science Master, nearly 30 years of industrial software experience, SCSAI PLM Chief Architect
- Agent Team: scsai (Chief Cross-border Officer) + exploreai (Founder of Transboundary Alliance)
Contact
- Business: Fu Bo (CEO) 18986066876 | fu@bossmind.com
- Technical: Fang Yunchao (CTO) 18601921816 | fang@scsai.com
- Experience: ylxt.chat (login-free access to all 71 digital employees)
