BossAgentsBossAgents × PLM Base · General Skill Integration Framework
BossAgents General Skill Integration Framework v1 (2026-08-31) — Build the underlying universal base once, then search and plug in verified external skills for upper-layer business scenarios. Procurement price comparison is merely the first instance, not a standalone system. Pairs with "Procurement Price Comparison Skill Integration Solution v1" and "Agent Performance Platform Detailed Design Solution v1".
| Metric | Value | Description |
|------|------|------|
| 1 set | General Base (shared by all scenarios) | Built once: 4-layer architecture for data assets / skill integration / runtime / scenario configuration |
| 5 steps | Integrate a new scenario | No base code modification: Select skills, configure mapping, set scoring, backtest, deploy |
| N | Pluggable scenarios | Procurement validated; maintenance/quality/process pending verification |
Contents: 01 Conclusion First · 02 Design Principles · 03 Overall Architecture · 04 Data Asset Registry · 05 Skill Registry · 06 Mapping Engine + Four Adapters · 07 Orchestration Runtime + Result Feedback · 08 Five Steps for New Scenario Integration · 09 Procurement Price Comparison Example · 10 Skeletons for Maintenance/Quality/Process Scenarios · 11 Implementation Timeline · 12 Risks and Countermeasures · 13 Data Source Description
01 Conclusion First
BossAgents requires not "one system for procurement, another for equipment maintenance," but a universal skill base built once and incrementally expanded.
In essence: The foundational four components (Data Asset Center, Skill Registry, Mapping & Adaptation, Orchestration Runtime) are universal engineering — built once and shared across all scenarios. Switching scenarios only requires three actions: select skills, configure mapping, set scoring — with zero base code changes.
As the first example, procurement price comparison validates not just "price comparison functionality," but also "end-to-end verification of this universal base in a real scenario." Once the base works, equipment maintenance, quality analysis, and process optimization follow the same five steps — only skills and mapping tables differ. This achieves "one digital employee with skills; switching employees requires no base rework."
02 Design Principles: Why a Universal Base is Essential
Comparison: one stack per scenario vs. common foundation + pluggable scenarios:
| Dimension | One stack per scenario (wrong approach) | Common foundation + pluggable scenarios (this framework) |
|------|--------------------------|------------------------------|
| Adding a new scenario | Rewrite mapping, adapters, and integration: 3-6 weeks/scenario | Select skill + fill mapping + configure scoring: 3-5 days/scenario |
| Foundation code | Maintain a separate set per scenario, gets messier over time | Maintain only one set shared by all scenarios, gets more stable with use |
| Data assets | Each scenario connects to databases independently; permissions/audit scattered | Unified data asset center with one permission model and one audit trail |
| Skill replacement | Changing a skill means modifying scenario code | Swap the skill implementation in the registry; fine-tune the mapping table |
| Delivery accumulation | Start from scratch each time; no experience accumulates | Each scenario delivered strengthens the foundation and skill library |
Foundation layer (common, built once):
- Data asset registry: all PLM data sources are registered and accessed uniformly
- Skill registry: skills are declaratively registered and invoked uniformly via MCP/API/SDK
- Data mapping engine: field mapping is configuration-driven; scenarios only fill in tables
- Four adapters (data/security/performance/output) are componentized and reusable
- Orchestration runtime: invocation, timeout, retry, audit, and result feedback
Scenario layer (pluggable on top):
- Each scenario only needs to: register data assets → select skills → fill mappings → configure scoring → validate
- Procurement comparison, equipment maintenance, quality analysis, process optimization... are all "scenario configuration packages"
- Scenarios are decoupled from each other; adding a new scenario does not touch existing ones
- Skills are integrated from externally searched, verified solutions (source code/MCP preferred), not self-developed
03 Overall Architecture: Common Foundation + Pluggable Upper Layer
Four-layer structure. The middle two layers form the common foundation (built once), the top layer is scenario configuration (pluggable), and the bottom layer is PLM data assets.
┌─────────────────────────────────────────────────────────┐
│ 场景层（插拔） │
│ 采购比价（样板）· 设备运维 · 质量分析 · 工艺优化 │
│ 每个场景 = 一个配置包 │
├─────────────────────────────────────────────────────────┤
│ 技能接入层（通用） │
│ 技能注册表 · 数据映射引擎 · 四项适配器 │
│ MCP / API / SDK 统一调用 │
├─────────────────────────────────────────────────────────┤
│ 运行时层（通用） │
│ 编排 · 鉴权 · 审计 · 重试 · 结果回流 │
│ 技能输出写回数据资产，形成闭环 │
├─────────────────────────────────────────────────────────┤
│ 数据资产层 │
│ PLM 统一注册：物料 / 历史价 / 供应商 / BOM / 设备台账 / │
│ 检测记录 / 工艺路线 │
└─────────────────────────────────────────────────────────┘
关键原则：场景只做「配置」，不做「开发」。技能不直接连 PLM 库，统一走数据资产中心取数；技能调用统一走注册表，任何技能可插拔替换。数据权限和审计在底座统一管控。
04 Foundation Component 1: Data Asset Registry
All PLM data sources are first registered as "data assets" in the foundation, with metadata (field name, type, unit, permissions). Scenarios no longer query tables directly via SQL; instead, they retrieve data through the unified data access interface get_data_asset(). This is done once and shared by all scenarios.
Example data asset registry (procurement and extended scenarios share the same registry):
| Asset ID | Data Source | Key Fields (Example) | Consumption Scenario | Status |
|---------|--------|------------------|----------|------|
| DA-001 | Material master data | Material code / specification / unit / category | Procurement price comparison, quality analysis | Registered |
| DA-002 | Historical purchase prices | Material / supplier / unit price / quantity / date | Procurement price comparison | Registered |
| DA-003 | Supplier profiles | Name / qualifications / rating / payment terms | Procurement price comparison | Registered |
| DA-004 | Purchase order history | PO number / on-time delivery rate / incoming material pass rate | Procurement price comparison | Registered |
| DA-005 | BOM structure | Product / component / bill of materials | Procurement price comparison, process optimization | Registered |
| DA-006 | Equipment ledger / operation logs | Equipment code / operating parameters / downtime records | Equipment maintenance (pending integration) | Pending registration |
| DA-007 | Inspection records / defective products | Batch / inspection item / defect type / process parameters | Quality analysis (pending integration) | Pending registration |
| DA-008 | Process routes / equipment parameters | Operation / parameter / energy consumption / output | Process optimization (pending integration) | Pending registration |
Each data asset carries a permission scope (which roles can view it) and access audit (who retrieved data and when)—done once, all scenarios remain compliant. Data assets can also flow back: new results generated by skills (price comparison conclusions, quality judgments) are written back to the asset pool, forming a closed data loop.
05 Foundation Component 2: Skill Registry
Each external skill is declaratively registered in the registry: ID, name, source repository, protocol type, input/output Schema, and verification status. The foundation invokes skills uniformly by protocol; scenarios do not need to care whether a skill is MCP or API.
Example skill registry (procurement instances are registered; the rest will be registered after search and verification):
| Skill ID | Name | Source | Protocol | Output Capability | Status |
|---------|------|------|------|----------|------|
| SK-PROC-01 | Procurement MCP Server | zavora-ai/mcp-procurement | MCP | price_benchmark / supplier_risk_score / supplier_recommend / negotiation_brief | Verified |
| SK-PROC-02 | WhichBid quote parsing | Dixon-O/which-bid | API | Quote parsing / 0-100 scoring / hidden fee detection | Verified |
| SK-PROC-03 | Price comparison scoring model | FastGPT practical weights | Algorithm | Composite score: price 40% + delivery 30% + quality 30% | Verified |
| SK-EQP-XX | Equipment fault prediction / maintenance planning | Pending search and verification | TBD | Fault warning / maintenance scheduling / spare parts prediction | Pending verification |
| SK-QA-XX | Quality root cause analysis / SPC | Pending search and verification | TBD | Abnormal root cause location / control charts / defect attribution | Pending verification |
| SK-PRO-XX | Process parameter optimization / energy consumption optimization | Pending search and verification | TBD | Parameter recommendation / energy comparison / yield optimization | Pending verification |
Structure of a skill registry entry (register new skills accordingly):
{
"skill_id": "SK-PROC-01",
"name": "Procurement MCP Server",
"source": "https://github.com/zavora-ai/mcp-procurement",
"protocol": "MCP",
"inputs": ["物料编码", "历史采购价", "供应商档案"],
"outputs": ["price_benchmark", "supplier_risk_score",
"supplier_recommend", "negotiation_brief"],
"license_check": "接入前核查许可证与依赖",
"status": "已验证（采购实例接入中）"
}
06 底座组件三：数据映射引擎 + 四项适配器
这是「换场景不换底座」的核心。字段映射做成配置化引擎，不是写死在代码里；场景接入只填一张映射表。四项适配器组件化实现一次，所有场景按需启用。
Mapping engine (scenarios only fill in the table; the engine is generic):
{
"场景": "采购比价",
"数据资产": ["DA-001物料", "DA-002历史价", "DA-003供应商", "DA-004订单", "DA-005BOM"],
"技能": ["whichbid.parse", "mcp.price_benchmark", "mcp.supplier_risk_score", "评分模型"],
"映射": [
{"源": "DA-001.物料编码", "目标": "whichbid.报价条目.物料编号"},
{"源": "DA-002.单价", "目标": "price_benchmark.历史均价"},
{"源": "DA-003.评级", "目标": "risk_score.资质得分"},
{"源": "DA-004.交期达成率", "目标": "评分模型.交付得分"},
{"源": "DA-004.来料合格率", "目标": "评分模型.质量得分"}
],
"编排": "解析报价单 → 比价基准 → 风险评分 → 综合排序 → 省钱报告",
"输出模板": "报价对比报告"
}
四项适配器：一次实现，场景配置化启用：
| Adapter | Common implementation (done once in the foundation) | What the scenario only does |
|--------|------------------------|--------------|
| ① Data adaptation | Field mapping engine, type conversion, unit unification, missing-value fallback | Fill in a mapping table (see above) |
| ② Security adaptation | Private deployment, sensitive-field masking, operation audit, data never leaves the factory | Check the sensitive fields involved in this scenario |
| ③ Performance adaptation | Connection pool, timeout, retry, cache, asynchronous task queue | Just configure the timeout threshold |
| ④ Output adaptation | Unified renderer: skill output → BossAgents artifact standard Schema | Select an output template |
Benefit: Because the adapters are generic, the "mapping engine + masking + audit + renderer" refined in the procurement scenario can be directly reused by equipment maintenance and quality analysis—no need to rewrite.
07 Base Component 4: Orchestration Runtime + Result Feedback
- Scenario orchestration: A scenario = an invocation sequence of a set of skills + scoring weights + fallback strategy, all configuration-based. Procurement orchestration is "parse → compare prices → risk → sort → report"; equipment maintenance is another sequence. Both are executed by the runtime according to configuration.
- Unified invocation: Dispatch according to the protocol in the skill registry (MCP client / REST / SDK); the scenario layer does not perceive the differences.
- Result feedback: After skill output passes through output adaptation, it is presented to digital employees and written back to the data asset center, serving as the historical basis for the next task.
- Authentication and audit: Every skill invocation records "who, when, which skill was called, and which data was used"—one unified set across all scenarios.
# 通用编排运行时（示意：一次实现，所有场景共用）
def run_scenario(scenario_config, request):
data = [data_asset.get(asset_id, filters) for asset_id in scenario_config["数据资产"]]
result = scenario_config["技能序列"](data) # 按注册表协议统一调用
report = output_adapter.render(scenario_config["输出模板"], result)
data_asset.write_back(report) # 写回数据资产，形成闭环
return report
08 上层插拔：新场景接入五步（不碰底座）
- 第 1 步 · 注册数据资产：从 PLM 选出该场景需要的数据源（设备台账、检测记录、工艺路线……），在资产中心登记字段与权限。
- 第 2 步 · 选技能 / 新增技能：优先从注册表复用；没有就用 AI 编程工具搜索该场景已被验证的最优技能（优先源码/MCP），验证后登记进注册表。
- 第 3 步 · 填映射表：把数据资产字段 ↔ 技能输入字段对应起来，写进场景配置包（JSON）。
- 第 4 步 · 配编排与评分：定义技能调用顺序、评分权重、输出模板、回退策略。
- 第 5 步 · 回测验证上线：用 PLM 历史数据回测，定验收指标（如比价准确率、故障预警提前率），达标后上线。
验收红线：任何场景，第 5 步没跑通历史数据回测，就不算接入完成。「演示能跑」不等于「数据喂进去结论靠谱」。
09 Scenario Example 1: Procurement Price Comparison (Showcase, Proven Solution)
Procurement price comparison is the first instance to fully complete the five steps, verifying that the base is usable. The detailed version of skill selection and field mapping is in the companion document "Procurement Price Comparison Skill Integration Plan v1"; this framework incorporates the distilled experience from the base perspective.
Procurement price comparison instance = a scenario configuration package (zero changes to base code):
| Integration Step | Approach in the Procurement Price Comparison Instance |
|----------|--------------------|
| ① Data assets | Materials / historical purchase prices / supplier profiles / PO history / BOM (DA-001~005) |
| ② Skill selection | Procurement MCP (price comparison/risk/recommendation/negotiation) + WhichBid (quote parsing) + scoring model |
| ③ Mapping Table | Material Code→Quotation Item, Historical Unit Price→Benchmark, Rating→Risk Score, Delivery Time/Qualification Rate→Scoring Weight |
| ④ Orchestration | Quotation Parsing → Price Benchmark → Supplier Risk → Comprehensive Ranking → Cost Savings Report |
| ⑤ Validation | Backtest with PLM historical procurement data, manually verify price comparison conclusions, complete first closed loop in 2 weeks |
After procurement validation, assets are solidified into the base platform: mapping engine logic, configurations of four adapters, output renderer, and cost savings report templates—all become directly reusable assets for subsequent scenarios.
10 Scenario Examples 2-4: Same base platform, different skills and configurations
The following three scenarios demonstrate 'switching scenarios without changing the base platform': data assets, skills, mappings, and orchestration are packaged as individual configurations, while the base platform's four components remain unchanged. Specific skill selections are searched/validated via AI programming tools and registered.
Configuration package skeletons for three pending integration scenarios (skill directions pending search/validation):
| Scenario | ① Data Assets | ② Candidate Skill Directions | ③④ Key Mapping & Orchestration Points | ⑤ Acceptance Criteria (Example) |
|------|------------|----------------|--------------------|--------------------|
| Equipment Maintenance | Equipment Ledger / Operation Logs / Inspection Records / Failure History | Failure Prediction, Maintenance Planning, Spare Parts Demand Prediction (pending search/validation) | Operation Parameters→Failure Feature Input; Orchestration: Monitoring→Alert→Scheduling→Work Order | Early Warning Rate for Failures, Reduction in Unplanned Downtime |
| Quality Analysis | Inspection Records / Defective Products / Process Parameters / Customer Complaints | SPC Control Charts, Root Cause Analysis, Defect Attribution (pending search/validation) | Inspection Items→SPC Input; Orchestration: Anomaly Detection→Attribution→Improvement Suggestions | Anomaly Detection Rate, Root Cause Localization Accuracy |
| Process Optimization | Process Routes / Equipment Parameters / Energy Consumption / Yield | Parameter Optimization, Energy Consumption Optimization, Yield Optimization (pending search/validation) | Process Parameters→Optimization Input; Orchestration: Baseline→Optimization→Parameter Recommendation | Energy Consumption per Unit Decrease, Yield Increase |
The skill directions and acceptance metrics in the table above are a skeleton to be validated, not finalized conclusions. Each scenario integration follows the five steps in Chapter 08, replacing skills with real solutions found via search, and setting metrics after backtesting with historical data.
11 Rollout Pace: Build the Foundation First, Then Insert Scenarios
- Weeks 1-2 · Build the Minimal Foundation Set: Data Asset Center (first register procurement-related DA-001~005) → Skill Registry Framework (unified MCP invocation) → Mapping Engine + four adapter skeletons. Goal: the foundation runs and can connect the first skill.
- Weeks 3-4 · Procurement Instance Fully Runs Through: WhichBid quote parsing + Procurement 2 core tools + scoring model, completing the five steps, backtesting with historical data, and producing a savings report template. This is the key step where "the foundation is validated by a real scenario."
- From Week 5 · Scenario Replication: Equipment O&M → Quality Analysis → Process Optimization, each integrated in 3-5 days. First search and validate the skill, then fill in the configuration package; the foundation code remains unchanged.
- Ongoing · Skill Library Accumulation: Each validated skill is deposited into the registry, and scenario configuration packages are archived. The value of BossAgents shifts from "many digital employees" to "thick skill library, stable foundation, fast scenario integration."
Reminder: Do not roll out scenarios first. If the foundation is not running and you integrate multiple scenarios in parallel, you are back to "one scenario, one set of code." Let procurement force the foundation to mature first, then replicate.
12 Risks and Countermeasures
Main risks and responses for building the general skill foundation:
| Risk | Level | Countermeasure |
|------|------|------|
| Foundation over-designed, trying to cover all scenarios in the first version | High | Prioritize the minimal foundation set: only build what supports the procurement instance running through, then expand based on scenario feedback |
| External skill quality varies / license / security risks | Medium | Only connect "validated" skills; check licenses and dependencies before integration; run in a private sandbox |
| Mapping table fields do not match, rework needed | Medium | Review the mapping table before development; archive as a configuration package; use fallback rules + confidence annotations for missing fields |
| Scoring weights do not match industry reality | Medium | Make weights configurable; calibrate via regression with PLM historical data |
| Data asset permissions/calibers are inconsistent across scenarios | Medium | Register fields and permissions uniformly in the asset center with one audit trail; handle caliber differences at the mapping layer as a fallback |
| Trying to integrate all skills at once | Low | Clarify "foundation first, then procurement, then replication" — integrating more does not mean stronger implementation |
13 Data Source Notes
The procurement instance skills in this framework come from public online search (as of 2026-08-31), serving as validated skills for the foundation's first instance:
- Procurement MCP Server (37 MCP tools, source-to-pay full process, 10 intelligent analysis tools): GitHub · Procurement MCP Server (zavora-ai/mcp-procurement)
- WhichBid (AI quote comparison, 0-100 scoring, hidden fee detection): GitHub · WhichBid (Dixon-O/which-bid)
- TenderTally (bid line-by-line Excel comparison, Claude skill, demo material): GitHub · TenderTally
- FastGPT Smart Procurement Practice (price 40% + delivery 30% + quality 30% weights, manufacturing cost reduction 12.7% case): CSDN · FastGPT Supplier Automatic Evaluation and Dynamic Price Comparison
Skill selection for equipment O&M / quality analysis / process optimization are directions to be validated; when integrating, use AI programming tools to search for real solutions and replace. Specific PLM field names need to be confirmed with the implementation team; person-days/cycles are engineering estimates, subject to actual development scheduling. See the companion document "Procurement Price Comparison Skill Integration Plan v1" for the detailed procurement plan.
BossAgents General Skill Integration Framework v1 ｜ BossAgents Project Team ｜ 2026-08-31 ｜ Companion: Procurement Price Comparison Skill Integration Plan v1, Agent Performance Platform Detailed Design Plan v1
