案例研究
Enterprise OA Agent
为 2,000+ 员工重构 OA 高频流程，把自然语言请求转化为可审批、可追踪、可回滚的系统动作。
系统快照
架构、信号与证据会先于正文呈现。- Users
- 2000+
- Form time
- 10s
- Efficiency
- 20x
Problem
企业 OA 系统中的流程入口分散，2,000+ 员工需要理解多个系统页面、审批规则和字段约束。单纯把模型接入聊天窗口并不能解决问题，真正困难的是把自然语言意图转成可执行、可审批、可审计的业务动作。
Constraints
- 工具调用必须可审计，不能绕过原有权限系统。
- 敏感操作需要审批和二次确认。
- Agent 执行失败时要能恢复现场，而不是只返回一段解释。
System Design
系统把用户输入拆成意图识别、参数补全、审批检查、工具调用和结果回写几个阶段。基于 LangGraph、FastAPI 与 FastMCP 构建 OA 智能体和 MCP 服务，打通待办、排班、请假、外勤、会议室预定等工具链。
知识库侧搭建采集、预处理、分块、嵌入和检索链路，并基于 MinerU 完成版面分析、OCR、公式识别等端到端解析；检索侧结合 RAG-Fusion 和 HyDE 提升召回与准确率。
Key Decisions
最关键的判断是让 Agent 服从流程，而不是让流程迁就 Agent。每个工具都需要幂等约束、权限检查和结构化返回，避免模型把一次失败解释成另一次行动。
Impact
- 表单处理时间由 2-3 分钟降至约 10 秒，效率提升近 20x。
- 项目验证了自然语言入口在流程密集型场景中的可行性。
- 沉淀出工具注册、审批拦截、执行日志和知识库链路等可复用能力。
Reflection
下一版会进一步增强评测集、异常恢复和跨系统状态同步，让 Agent Runtime 更接近生产级工作流引擎。
Problem
Enterprise OA workflows are scattered across many systems. More than 2,000 employees need to understand pages, approval rules, and field constraints before they can act.
Adding a model to a chat window is not enough. The harder part is translating natural language into executable, approvable, and auditable business actions.
Constraints
Tool calls must be auditable and must not bypass existing permission systems.
Sensitive operations need approval and explicit confirmation.
When the Agent fails, the system needs recoverable state rather than a plain explanation.
System Design
The system separates user input into intent recognition, parameter completion, approval checks, tool execution, and result write-back.
LangGraph, FastAPI, and FastMCP connect OA agents with workflow tools, while MinerU supports document parsing for internal knowledge flows.
Key Decisions
The Agent should obey the workflow rather than forcing the workflow to adapt to the Agent.
Each tool needs idempotency, permission checks, and structured returns so a failed action is not turned into another unsafe action.
Impact
Form processing time dropped from 2-3 minutes to roughly 10 seconds, creating a reusable intelligent-office blueprint.
Reflection
The next version should strengthen evaluation sets, exception recovery, and cross-system state synchronization so the Agent Runtime behaves more like a production workflow engine.