An AI agent is a system built on top of a large language model (LLM) that can plan multi-step tasks, decide when to call external tools (search, APIs, code execution, file I/O), observe the results, and loop until the task is complete. Unlike a chatbot that just responds, an agent decides what to do next.

How does an AI agent actually work?

An agent runs a loop: 1) receive the user's request; 2) send it to the LLM to decide the next action; 3) if a tool is needed, call it and collect the result; 4) feed the result back to the LLM; 5) repeat until the LLM decides it has enough to answer; 6) return the final answer to the user. This is often called the think-act-observe or ReAct loop.

What's the difference between an AI agent and a chatbot?

A chatbot replies to a message with another message using only what the LLM already knows. An AI agent can call tools to fetch fresh data, run code, read files, or take actions in the real world, and it plans multiple steps before responding — turning the LLM from a chat responder into a task executor.

What are the core components of an AI agent?

A typical agent has: (1) a system prompt that defines its role; (2) tool definitions (JSON schemas the LLM can call); (3) an LLM for reasoning; (4) a function-calling mechanism to route tool calls; (5) short and long-term memory; (6) a context window that limits how much history it can see.

What is function calling in AI agents?

Function calling is the mechanism by which an LLM emits a structured request to invoke an external function instead of replying with plain text. The agent runtime intercepts that request, executes the actual function, and feeds the result back into the LLM so it can continue reasoning.

🤖 How AI Agents Work · Interactive Animation

How do AI Agents actually work?

Watch an interactive animation of an AI agent's full reasoning loop:
user question → LLM decides → tool call → observed result → final answer.

🎬

See the Animation

Watch an agent solve a task, step by step

🧠

Core Concepts

LLM, function calling, tools, memory

👤

→

🤖

→

🧠

🎬 How AI Agents Work — Visualized

The clearest way to see what actually happens inside an AI agent

An AI agent is a system built on top of a large language model that can call external tools, plan multi-step actions, and loop until a task is done. It doesn't just answer — it decides what to do next, invokes the right tool (search, an API, code, a file), observes the result, and iterates.

The animation below walks through the whole loop: user asks → agent consults the LLM → tool is discovered & called → result is fed back → final answer is generated.

🧠 Think

→

🔧 Use Tools

→

✅ Answer

🎬 Interactive Animation: An Agent's Full Loop

Pick a scenario and watch every step — from the user's message to the final reply. Adjustable speed · step-by-step mode.

0/0

What can AI agents do?

💻

Write code

Read requirements, write code, run tests, fix bugs — a full-stack pair programmer that never gets tired.

🔍

Research the web

Search, read, cross-check sources, and summarize the findings — so you don't have to open 30 tabs.

📊

Analyze data

Query databases, generate charts, write reports — automated analysis pipelines from a single prompt.

🎫

Handle support

Understand a customer's issue, look up their order, issue refunds, escalate the hard cases — 24/7.

📅

Book & buy

Search flights, compare prices, fill forms, complete checkouts — agents can take real actions, not just talk.

📝

Write & edit

Draft docs, translate content, polish tone, rewrite for clarity — a writing collaborator on call.

AI Agent vs. traditional app

Dimension	📱 Traditional app	🤖 AI agent
How it works	Fixed flow, buttons trigger preset actions	Understands intent, plans steps, picks tools dynamically
Understanding	Only handles predefined input formats	Understands natural language and fuzzy requests
Tool use	Features hard-coded, hard to extend	Discovers and calls new tools on demand
Adaptability	New scenario ⇒ ship a new release	Can reason about situations it has never seen
Example	Weather app → just shows weather	Agent → "check weather + suggest what to wear + plan the trip" in one shot

🧠 Core Concepts of an AI Agent

The 6 building blocks that turn an LLM into an agent

Agent Architecture Overview

The 6 core concepts

📝

System Prompt

The system prompt defines the agent's role, capabilities, and behavior boundaries — its "persona."

              // System prompt for a weather assistant

              You are a professional weather assistant.

              When the user asks about weather,

              you must first call the get_weather tool

              to fetch real-time data, then give clothing advice.

              Never invent weather data you did not fetch.

🔧

Tool Definition

Tool definitions tell the LLM what external capabilities exist and what parameters each one takes.

function get_weather({
  "city": "string",     // e.g. "Tokyo"
  "date": "string"      // optional, ISO date
}) returns {
  "temp": "number",    // °C
  "condition": "string",// e.g. "sunny"
  "humidity": "number"  // percent
}
            

🔗

Function Calling

The LLM doesn't run tools directly — it emits a structured "call this function with these args" request that the agent runtime executes.

// LLM output is a call, not prose
{
  "tool_call": "get_weather",
  "arguments": {
    "city": "Tokyo",
    "date": "2026-07-03"
  }
}
// The runtime picks it up and executes.
            

🪟

Context Window

The context window is the maximum amount of text the LLM can "see" at once — it caps how much history the agent can remember.

Windows vary widely by model:
• Early models: 4K–8K tokens
• Modern models: 128K–200K tokens
• Newest models: 1M+ tokens

1 token ≈ 0.75 English words. Anything beyond the window is "forgotten," so agents need smart memory management.

🔄

Multi-turn Reasoning

Agents don't answer in one shot. They loop through think → act → observe → think again until they can respond.

A typical reasoning loop:
1️⃣ User asks a question
2️⃣ Agent → LLM: which tool should I use?
3️⃣ LLM → Agent: use get_weather
4️⃣ Agent calls the tool, gets a result
5️⃣ Agent → LLM: given this data, how do I reply?
6️⃣ LLM → Agent: final response
7️⃣ Agent → User: the answer

🧠

Memory & State

Agents need memory to stay coherent — short-term for the current conversation, long-term for user preferences and past sessions.

Short-term memory: current conversation, lives in the context window.
Long-term memory: facts across sessions, stored in an external DB or vector store.
Working memory: intermediate results while a task is in progress.

Good memory design is what lets an agent handle long, complex tasks.

Why understand how AI agents work?

🎯

Prompt more precisely

Knowing when the agent is thinking vs. calling a tool lets you write prompts that get it right on the first try.

🛡️

Trust it appropriately

Understanding context windows, hallucinations, and tool dependencies tells you what to verify — and what to leave alone.

🚀

Build your own

Going from user to builder starts here. The concepts on this page are the foundation for every agent framework out there.