From Agent to Skills: A Paradigm Shift in AI Agent Architecture

技术报告：从 Agent 到 Skills — AI 智能体架构的范式转变

This repository contains the companion code and technical report for understanding the paradigm shift from monolithic AI Agents to composable Skills + MCP + A2A architecture.

What’s Inside

Technical Report

• 技术报告_从Agent到Skills的范式转变.md — Full technical report (Chinese), covering:
  • Why Skills replace monolithic Agents
  • Core concepts: Agent, Skills, MCP, OpenClaw, A2A
  • Real-world project walkthrough
  • Experiment: Claude Code vs OpenClaw side-by-side comparison

Skills (Working Code)

skills/
├── auto-twitter-campaign/     ← Core project: AI image processing + tweet generation
│   ├── SKILL.md               ← Skill metadata + instructions (3-layer progressive disclosure)
│   └── scripts/
│       ├── campaign.py        ← Orchestrator (thin layer)
│       └── batch_compare.py   ← Image pipeline (Director-Creator-Critic pattern)
│
├── wechat/                    ← WeChat keyboard automation skill
│   ├── skill.json + prompt.md
│   └── scripts/               ← Shell wrappers for send/read
│
└── wechat-bridge/             ← WeChat ↔ OpenClaw bridge
    ├── skill.json + prompt.md
    └── scripts/
        ├── demo_case1.py      ← Claude Code approach (deterministic script)
        ├── bridge_openclaw.py ← OpenClaw approach (thin IO adapter + Agent reasoning)
        └── bridge.py          ← Original version (superseded)

Key Architecture

┌─────────────────────────────────────────────────────┐
│                  2026 Agentic AI Stack               │
│                                                     │
│  Application:  OpenClaw / Claude Code / Cursor       │
│  Knowledge:    Skills (teach HOW to do things)       │
│  Reasoning:    Agent (decide WHAT to do)             │
│  Tools:        MCP (connect to external services)    │
│  Collaboration: A2A (Agent-to-Agent delegation)      │
│  Foundation:   LLM API / Local models                │
└─────────────────────────────────────────────────────┘

Requirements

• Python 3.10+
• google-genai SDK
• Pillow for image processing
• ZENMUX_API_KEY environment variable (for Ming-flash-omni-2.0 API)
• OpenClaw (optional, for Agent-based execution)

License

MIT