Lecture

Agents

15 Sep 2025

Intro

The Birth of An Agent: Construction of LLM-based Agents

AI Agents Architecture

Paper: AI Agents: Evolution, Architecture, and Real-World Applications

AI Agent 擬人化角色與分工

NPCpy

Code: https://github.com/NPC-Worldwide/npcpy

from npcpy.npc_compiler import NPC
simon = NPC(
          name='Simon Bolivar',
          primary_directive='Liberate South America from the Spanish Royalists.',
          model='gemma3:4b',
          provider='ollama'
          )
response = simon.get_llm_response("What is the most important territory to retain in the Andes mountains?")
print(response['response'])

Frameworks

OpenAI Agent SDK

Code: https://github.com/openai/openai-agents-python
pip install openai-agents

AG2: Open-Source AgentOS for AI Agents

Examples:

OpenEvolve

Code: https://github.com/algorithmicsuperintelligence/openevolve
Turn your LLMs into autonomous code optimizers that discover breakthrough algorithms

ShinkaEvolve

Paper: ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution
Code: https://github.com/SakanaAI/ShinkaEvolve

AlphaResearch

Paper: AlphaResearch: Accelerating New Algorithm Discovery with Language Models
Code: https://github.com/alpharesearchbot/Alpha

GSW (Generative Semantic Workspaces)

Paper: Beyond Fact Retrieval: Episodic Memory for RAG with Generative Semantic Workspaces

Agentic Coding

Claude Code

Install Claude Code curl -fsSL https://claude.ai/install.sh | bash

OpenCode

Install: curl -fsSL https://opencode.ai/install | bash

OpenCode setup: Beginner’s Crash course

AlphaResearch

Paper: AlphaResearch: Accelerating New Algorithm Discovery with Language Models
Code: https://github.com/alpharesearchbot/Alpha
](https://github.com/anomalyco/opencode)

# YOLO
curl -fsSL https://opencode.ai/install | bash

# Package managers
npm i -g opencode-ai@latest        # or bun/pnpm/yarn

Antigravity

Blog: Introducing Google Antigravity
「反重力」系統它內部有好幾個AI代理（Agent），就像一個項目團隊：

「規劃師」代理：負責把你的模糊目標，拆解成具體的、可執行的步驟。
「研究員」代理：負責上網查找最新的天氣API接口是什麼，最好的UI設計長什麼樣。
「構建師」代理：負責根據規劃和研究，吭哧吭哧地寫代碼。
「測試員」代理：負責檢查代碼有沒有bug。
「反思者」代理：負責評估整個過程，看看哪裡可以改進。

MCP - Model Context Protocol

MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems.
Blog: Standardizing AI Tooling with Model Context Protocol (MCP)

MCP Servers

Connect Claude Code to tools via MCP

FreeCAD-MCP

Solidworks-MCP

**https://github.com/vespo92/SolidworksMCP-TS

Skills

Agent Skills

The Skills architecture

my-skill/
├── SKILL.md          # Required: instructions + metadata
├── scripts/          # Optional: executable code
├── references/       # Optional: documentation
└── assets/           # Optional: templates, resources

https://github.com/anthropics/skills

NotebookLM-skill

Install:

mkdir -p ~/.claude/skills
cd ~/.claude/skills
git clone https://github.com/PleasePrompto/notebooklm-skill notebooklm

Run agentic coding tool:

claude or opencode

Inputs:

What are my skills?

Set up NotebookLM authentication

Add this NotebookLM to my library : https://notebooklm.google.com/notebook/c2e2c6de-0131-420e-8686-2326bc9a438a

What is Agent Skills ?

UIUX-Pro-Max-skill

Install:

npm install -g uipro-cli

uipro init --ai claude # Claude Code
uipro init --ai gemini # Gemini
uipro init --ai trae # Trae
uipro init --ai opencode # OpenCode

Go to project:

cd ~/GenAI/Skills/notebooklm

Run agentic coding tool:

claude or opencode

Inputs:

Build a landing page for my SaaS product

OpenAI Symphony

Symphony turns project work into isolated, autonomous implementation runs, allowing teams to manage work instead of supervising coding agents.

Claw Family

Clawra

Openclaw as companion

OpenClaw Architecture

Blog: OpenClaw (Clawdbot) Architecture: Engineering Reliable and Controllable AI Agents

The 6-Stage Execution Pipeline:

Channel Adapter: Standardizes inputs from different platforms (e.g., Discord or Telegram) into a unified message format while extracting necessary attachments.
Gateway Server: Acts as a session coordinator, determining which session a message belongs to and assigning it to the appropriate queue.
Lane Queue: A critical reliability layer that enforces serial execution by default, allowing parallelism only for explicitly marked low-risk tasks.
Agent Runner: The “assembly line” for the model. It handles model selection, API key cooling, prompt assembly, and context window management.
Agentic Loop: The iterative cycle where the model proposes a tool call, the system executes it, the result is backfilled, and the loop continues until a resolution is reached or limits are hit.
Response Path: Streams final content back to the user channel while simultaneously writing the entire process to a JSONL transcript for auditing and replay.