Introduction to AI Agent
Last updated
Last updated
Why is everyone suddenly talking about AI Agents?
In this section, we’ll explore an AI Agent, how it relates to large language models (LLMs), why structured data matters, and how to design an Agentic Workflow, with images and examples for clarity.
An AI Agent is a system that can understand user instructions or contextual situations and autonomously execute tasks to achieve a given goal.
To do so, it can leverage various tools and APIs, making decisions based on a data-driven understanding of the situation.
✅ In short, an AI Agent is an intelligent system that acts autonomously to accomplish a given objective.
Large Language Models (LLMs) are primarily designed for understanding and generating text.
As LLMs have grown in scale, their reasoning capabilities have significantly improved—this is the power of the Scaling Law.
However, solving real-world problems requires more than just generation.
That’s where AI Agents come into play—expanding LLMs' capabilities by integrating additional tools and workflows.
Agents take advantage of LLMs' reasoning abilities and combine them with other capabilities, like computation, web search, or external API calls, to automate complex tasks.
One of the key technologies that makes this possible is Function Calling.
Want to dive deeper?
👉 Check out "Getting Started with Function Calling" for more details!
Function Calling
allows an LLM to "request" the use of an external tool. You can think of it like an AI pressing a button in a program to act.
For example:
When asked: “What’s the weather like in Seoul tomorrow?”
→ The LLM can respond with something like:
get_weather(location="Seoul", date="tomorrow")
@tool decorator: Registers this function as a callable tool.
Docstring: Describes the function’s purpose—this helps the LLM decide when to use it.
So, LLMs can understand a user’s intent and express it as a function call, like a developer writing code.
Here's how LLMs and Agents work together:
LLM analyzes the question and generates a suitable function call
The agent executes the actual function
LLM receives the result and presents it in natural language
Summary:
Function Call = A request from the LLM to use an external tool (API, calculator, search engine, etc.)
Agent = The assistant that executes the request and returns the result
For multiple agents to collaborate, they can't just rely on ambiguous natural language.
❌ Unstructured: “아버지가방에들어가신다” (no spaces = unclear)
✅ Structured: {"Person": "Father", "Place": "Room"}
Agents collaborate effectively using structured data formats like JSON.
Want to learn more?
👉 Check out "Getting Started with Structured Output" for deeper insights!
How can we design a practical Agentic Workflow?
An Agentic Workflow is a structured design approach that breaks down a complex goal into smaller tasks and assigns each task to the most suitable Agent to handle it.
🔍 The key is task decomposition.
It’s important to note that task decomposition must be done by humans—not the LLM.
LLMs or Agents alone can't fully understand how to break down complex goals. We, as designers, need to understand the objective and logically deconstruct it into manageable steps.
Only then can we effectively assign each task to a specific agent.
Let’s look at a few examples:
User request: “Do I need to bring an umbrella in Seoul tomorrow?”
🧩 Task Breakdown:
Identify location and date
→ Seoul, tomorrow
Retrieve weather data
→ Check the chance of rain
Apply rain threshold logic
→ Recommend an umbrella if the chance ≥ 60%
Generate a natural language response
🤖 Example Agent Assignment:
NLP Agent: Extracts date and location from the query
Weather Agent: Calls the weather API
Judge Agent: Applies decision rules (e.g., bring an umbrella if needed)
Response Agent: Generates a user-friendly response
User request: “Create a promotional email for our new product.”
🧩 Task Breakdown:
Analyze target customers
Summarize product details
Choose email tone/style
Write email content
Add personalized suggestions
🤖 Example Agent Assignment:
Customer Analysis Agent: Uses CRM data to identify customer segments
Product Summary Agent: Extracts product highlights
Tone Styling Agent: Adapts style to marketing goals
Copywriting Agent: Writes persuasive email content
Personalization Agent: Adds behavior-based recommendations
Data Analysis Agent: Analyzes large CSV files, visualizes data, and fixes errors
File Management Agent: Automates file deletion, movement, and reading/writing
Report Writing Agent: Searches documents, summarizes them, and creates formatted reports with images
Agentic Workflow = Task Decomposition + Role Assignment
Humans must define the tasks strategically
Each Agent handles what it does best
Overall structure: Router → Task → Agent → Response
As we’ve seen, designing workflows where multiple Agents collaborate is essential. To make this collaboration efficient, we need a smart controller:
The Orchestration Layer.
The orchestration layer acts like a conductor, coordinating the right Agents or Models based on the user's input.
It understands the user’s intent
Chooses the most efficient Agent combination
Assigns roles and sequences tasks appropriately
Modern AI systems no longer rely on a single model. Instead, they combine multiple Agents and Models with specialized abilities.
✅ Service Stability
→ If one Agent fails, others can still complete the task
✅ Performance Optimization
→ Select high-efficiency or low-cost models as needed
✅ Collective Intelligence
→ Multiple Agents working together yield better outcomes
Users express goals in many ways:
“Make it accurate.” “I need it cheap.” “Exclude models from Country X.”
“Speed is critical.” “I need the result today.”
Based on these constraints, the orchestration layer dynamically chooses the best model and agent combination.
Agents are becoming more intelligent and interactive.
They are evolving from simply executing commands to:
Recognizing what they can/can’t do
Requesting missing information
Making autonomous decisions based on context
In this section, we covered the core concepts behind AI Agents:
🔹 What is an AI Agent?
An autonomous system that reasons and acts to achieve a goal
🔹 LLM + Agent Architecture
LLMs handle reasoning; Agents handle external execution
🔹 Structured Data
Essential for Agent collaboration
🔹 Agentic Workflow
Break tasks down and assign them to the right Agent
📘 Up next, we’ll get hands-on and start building your very first AI Agent using Upstage APIs. Ready to create your own Agent? Let’s go! 🤖✨
YoungHoon Jeon | AI Edu | Upstage
