Seri Belajar LLM Part 5

LLM Agents & Tool Use

Dari chatbot ke autonomous agent yang bisa browse, search, code, dan email. Part 5 mengajarkan Function Calling, ReAct pattern, MCP protocol (standard 2025), multi-agent systems, dan framework populer (LangChain, CrewAI, OpenClaw).

Maret 202630 menit bacaAgents • Tool Use • Function Calling • MCP • Multi-Agent

📚 Seri Belajar LLM:

1 2 3 4 5 6 7 8 9 10

Daftar Isi

Chatbot vs Agent — Agent bisa bertindak, bukan hanya menjawab
Agent Loop — Observe, Think, Act, Repeat
Function Calling — LLM memutuskan tool mana yang dipanggil
MCP Protocol — Standard 2025 untuk LLM-to-tool connection
Multi-Agent Systems — Kolaborasi agent untuk task kompleks
Kode: Agent dengan Tools — OpenAI function calling demo
Frameworks — LangChain, CrewAI, AutoGen, OpenClaw
Ringkasan —

🤖

1. Dari Chatbot ke Agent

Chatbot menjawab pertanyaan. Agent BERTINDAK: browse, search, code, email, booking.

LLM biasa (chatbot) hanya bisa menerima teks dan menghasilkan teks. Agent adalah LLM yang dilengkapi dengan tools (web search, calculator, code execution, API calls, database queries) dan autonomy untuk memutuskan tool mana yang dipakai, kapan memakainya, dan bagaimana menggabungkan hasilnya. Agent bisa menyelesaikan tugas multi-step yang sebelumnya butuh manusia: riset kompetitor, analisis data, scheduling meeting, bahkan debugging kode.

Agent Loop — Observe, Think, Act, Repeat

🔧

2. Function Calling — LLM Memilih Tool

LLM menerima daftar tools (JSON schema), dan memutuskan kapan memanggilnya

10_function_calling.py

# Define available tools
tools = [{
    "type": "function",
    "function": {
        "name": "search_web",
        "description": "Search the web for current information",
        "parameters": {
            "type": "object",
            "properties": {
                "query": {"type": "string", "description": "Search query"}
            },
            "required": ["query"]
        }
    }
}]

response = client.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "What is the weather in Jakarta today?"}],
    tools=tools
)
# LLM returns: tool_call = search_web(query="Jakarta weather today")
# Your code executes the search, returns results to LLM
# LLM generates answer with real, current data

🔌

3. MCP — Model Context Protocol

Standard terbuka 2025 dari Anthropic untuk menghubungkan LLM ke tools dan data

Protocol	Creator	How It Works	Adoption
MCP	Anthropic (2024)	Standardized tool/data interface. JSON-RPC over stdio/HTTP	Claude, VS Code, 50+ integrations, joined Linux Foundation
Function Calling	OpenAI	Native API parameter for tool definitions	GPT-4, Assistants API
Tool Use	Anthropic API	Similar to function calling, built into Claude API	Claude API users
LangChain Tools	LangChain	Python abstraction for tool integration	Massive community

👥

4. Multi-Agent Systems

Beberapa agent berkolaborasi untuk task kompleks

Single agent bagus untuk tugas sederhana. Untuk tugas kompleks (riset, analisis, decision-making), multi-agent systems membagi tugas ke beberapa agent spesialis. Contoh: Agent Researcher (cari data), Agent Analyst (analisis data), Agent Writer (tulis laporan), Agent Reviewer (quality check). Framework: CrewAI, AutoGen, LangGraph, OpenClaw.

Framework	Focus	Strengths	Stars (GitHub)
LangChain/LangGraph	General purpose agent orchestration	Huge ecosystem, flexible	95K+
CrewAI	Role-based multi-agent collaboration	Easy to use, good defaults	45K+
AutoGen (Microsoft)	Multi-agent conversation patterns	Research-grade, Microsoft backed	38K+
OpenClaw (OpenManus)	Open-source computer-use agent	Built-in browser, file system, tools	180K+
Semantic Kernel (Microsoft)	Enterprise AI orchestration	C#/.NET focus, enterprise ready	23K+

📙

Next: Part 6 — Local LLMs

Jalankan LLM di laptop Anda. Ollama, llama.cpp, quantization GGUF. Privacy, gratis, offline.

LLM

Tech Review Desk — Seri Belajar LLM

Sumber: Sebastian Raschka, Anthropic, OpenAI, Hugging Face, LLMOrbit, DeepSeek technical reports.

rominur@gmail.com • t.me/Jekardah_AI