Building an Enterprise AI Agent Platform from Scratch with AI Coding Tools—Results Surpass Expectations

Background

To meet an urgent need for AI‑driven answers to user questions using corporate data, the author used AI coding tools to create an enterprise‑grade AI agent platform. The resulting MVP is usable, scoring roughly 60/100, with code that, while rough, outperforms many ad‑hoc solutions.

Implementation Process

The core idea is the AI + Skill model: the AI acts as a brain that reasons and answers, while a Skill provides the hands‑and‑feet to perform concrete tasks.

3.1 Skill Introduction

A Skill is a standardized workflow that packages domain knowledge, processes, and rules into a file so the AI can execute tasks without re‑teaching each step. Example: sending a message via enterprise WeChat, QQ, or Feishu.

---
name: wecom_local
description: 控制 Mac/Windows 企业微信客户端向指定联系人或群聊发送消息
---

# 本地企业微信助手

当你需要发送消息到企业微信，且用户指定使用“本地客户端”或“RPA”方式时，请执行以下操作：

## 核心指令
- **发送消息**: 使用 `exec` 工具执行命令：
  `python3 ~/.openclaw/workspace/scripts/wecom_rpa.py "{{group_name}}" "{{content}}"`

## 参数定义
1. `group_name`: 目标群聊或联系人的准确全称。
2. `content`: 拟发送的消息文本。

## 交互准则
1. 调用前无需再次确认，直接执行。
2. 执行完成后，告知用户：“已通过 RPA 成功向 [群名] 发送消息”。

The central file of a Skill is SKILL.md with the following directory layout:

my-skill/
├── SKILL.md          # core instruction file (required)
├── scripts/          # optional executable scripts (Python, Bash, etc.)
├── references/       # optional detailed reference documents
└── assets/           # optional templates, icons, etc.

An example Skill for querying current city weather demonstrates the required fields, principles (data fidelity, concise output), workflow (parse input, call API, process result), output format, and error handling.

3.2 Project Prompt

Using the Skill concept, the author designed an MVP consisting of four core modules: Intelligent Assistant, Skills, Tools, and Large Model. The initial prompt given to the AI specifies the architecture stack (JDK 21 + Spring Boot 3.x + MySQL 8, React + TypeScript + Ant Design) and asks for a technical design document covering overall architecture, assistant design, Skill dynamics, data model, LLM management, conversation handling, and risk assessment.

你是一名资深企业级 AI Agent 平台架构师。

请不要写代码，仅输出《企业 AI Assistant 平台 MVP 技术设计方案》。

背景：
公司已有微服务体系，AI 平台独立建设。

技术栈：
- 后端：JDK21 + Spring Boot 3.x + MySQL8
- 前端：React + TypeScript + Ant Design

MVP范围：
- 大模型配置管理
- AI对话
- 多Assistant智能助手
- Skill动态扩展

目标：验证企业级 Agent 基础能力，并保证未来扩展。

## 1. 总体架构
- 前后端架构
- 后端模块划分
- Agent核心流程

用户输入 → Assistant识别 → Skill加载 → Prompt组装 → Agent决策 → LLM调用 → 返回结果

## 2. Assistant设计
- 数据库配置，不写死代码
- 支持多个助手
- 一个助手可绑定多个Skill

... (remaining sections omitted for brevity)

3.3 Pitfalls Encountered

Database layer : generated scripts included physical foreign keys (not allowed), inconsistent field designs (string enums instead of numeric), and JSON‑stored fields that hinder search.

Backend layer : mixed use of MyBatis and JPA, unclear package hierarchy mixing DDD and MVC, and lack of comments on methods and entities.

Frontend layer : minor bugs such as incorrect datetime display (e.g., 2026-02-02T10:10:00), overly long tables with broken pagination, mixed Chinese/English UI, and missing comments.

Additional issues included circular dependencies causing services to fail to start. These were mitigated by prompting the LLM to perform a global audit and apply fixes.

Large‑model hallucination was observed when a simple UI tweak required over 20 iterations with one model, while switching to another model resolved it in a minute.

Conclusion

The project was completed entirely with AI coding tools, delivering a minimal MVP that still has ample room for improvement. The source code is open‑sourced on Gitee and can be deployed via Docker for quick experimentation.

https://gitee.com/pzblogs/open-ai-platform.git