AI Agents, MCP & Function Calling – Visual Notes

🤖 AI Agents, MCP & Function Calling

A clear breakdown of system prompt, function calling, agent tools, and MCP — without the jargon.

💬 The Question

Agent, MCP, function calling — what do they actually mean? Is MCP really the “USB protocol for AI”? And what is an agent? Let’s connect all the dots in plain English.

📝 1. User Prompt vs. System Prompt

In early 2023, an AI was just a chat box. You send a user prompt — it replies. But without a persona, answers felt flat. “My stomach hurts.” “Go to the hospital.” Boring.

Then we started adding system prompts — the AI’s role, tone, background. “You are my girlfriend.” Now “My stomach hurts” → “Get lost, my stomach hurts too.” It felt right.

System prompt is metadata (personality, rules). User prompt is the actual question. They get merged before the model sees them.

ChatGPT’s “Customize ChatGPT” feature lets you set system prompt preferences.

⚡ 2. From Chatbot to Agent: AutoGPT

open source AutoGPT was a pioneering agent. You register functions (e.g., list_files, read_file) with descriptions. The system prompt tells the AI what tools exist and how to call them.

The AI returns a formatted request. AutoGPT parses it, executes the function, feeds back the result — loops until the task is done. This “middleman” program is called an AI agent, and the functions are agent tools.

Problem: AI sometimes returns wrong formats → retry loops (still used by many agents like Kling).

🔧 3. Function Calling: Standardization

To reduce flaky retries, big model makers (OpenAI, Google, Anthropic) introduced function calling.

Each tool is described via a JSON object (name, description, parameters). These JSONs are moved out of system prompt into a separate field. The AI’s response format is also fixed. If the model replies incorrectly, the server can retry seamlessly — the user doesn’t notice.

Unified format → easier training
Less token waste on client side
Still, no universal standard across vendors → both system prompt & function calling coexist.

🌐 4. MCP: Model Context Protocol

Tools like “web browse” are reused across many agents. Copying code is messy. So we wrap tools as services — that’s MCP.

MCP is a communication protocol between agent (MCP client) and tool service (MCP server). It defines:

How to discover tools, resources, and prompt templates
Can run locally (stdin/stdout) or remotely (HTTP)
MCP itself is model-agnostic — it only manages tools, not which AI model is used.

🔄 5. How They All Fit Together

Example: “My girlfriend has a stomach ache, what should I do?”

User asks the agent (MCP client).

Agent fetches tool info from MCP server (e.g., web_browse).

Agent converts tool info into system prompt or function calling format, plus user prompt, and sends to the AI model.

AI model decides to use web_browse to search online. Returns a function call request.

Agent calls MCP server’s web_browse, gets content, relays back to AI.

AI synthesises answer → “Drink more warm water.” Agent shows it to the user. She’s impressed 😊

System prompt gives the AI a persona.
Function calling standardizes tool descriptions.
Agent orchestrates the flow.
MCP connects agents to tools like a universal plug.

They are not replacements — they work together like gears in a system.

🌱 Final Reflection

“Many fear AI will replace us. I don’t know about the future, but I feel something else — a quiet thrill. We’ve always been pushed along by technological waves. This time, I want to meet it awake. Maybe it starts with understanding a small AI agent.”

— A programmer’s note

Generated by CuteNote.app Visual Notes

AI Agents, MCP &amp; Function Calling – Visual Notes

AI Agents, MCP & Function Calling – Visual Notes