Author(s): Nileka Samarasinghe

Originally published on Towards AI.

From not knowing what is an AI Agent to building Multi Agent AI Systems

You open twitter (or X, whatever) and suddenly everyone is shouting this fancy words at you

‘agentic ai’, ‘large action models’, ’autonomous orchestration’, ‘agent swarms’.

okay hold up. WTH is this?

If you’re confused, don’t worry. I was too. But after days of deep-diving, struggling through code, and finally building my own multi agent system, I’ve cracked the code. Here is what I actually learned about AI Agents.

First off what is an LLM, Anyway?

Before we run, we have to walk.

We have to talk about Large Language Models(LLMs).
Think of an LLM like a contestant on a trivia game show who has memorized the entire internet up until last year. They have read billions of books, articles, and websites. They are incredibly smart at predicting what word comes next in a sequence.
But here is the catch.
If you ask this trivia genius “ What is currently in my fridge? ” or “ Who won the game 5 min ago? ”, they will look at you blankly. They don’t know you, and they don’t know the now.

So examples of LLMs include GPT, Gemini, Claude, and LLaMA, while applications like ChatGPT, Gemini and Claude are end-user products built on top of these models.

You can actually see this happen if you treat ChatGPT like a best friend who knows your drama.
You: “So…… do you think he was lying? ”
ChatGPT: “Lying is a complex psychological behavior. Who is ‘he’? ”
You: “Dave! From the party last night! The one with the bad haircut!”
ChatGPT: “I do not know Dave. I was not invited to the party. Also, I have no eyes to judge haircuts.”
see the problem? It’s a genius at general knowledge, but it’s completely useless at knowing your context or what happened 5 minutes ago.

So, if an LLM is just a really smart brain, how do we make it actually do stuff?

That’s when AI Workflows come in.

This is where we start connecting the brain to the outside world. An AI Workflow is when we give the LLM a specific, pre-defined path to follow. It’s like a train on a track. It can only go where the rails are laid.

For example, imagine you connect an LLM to your Google Calendar. You ask, “When is my math test?” The system looks at your calendar and says, “March 8th.” Great!.

BUT — if you follow up with a question like, “Will it rain during the test? ”, the system crashes 🙁
Why? because you didn’t build a track for the Weather App. It only knows the path to the Calendar.
It’s rigid. It can’t pivot.

Wait, isn’t this what people call RAG(Retrieval-Augmented Generation)?

Yep! but RAG is just one flavor of the Workflow ice cream. There are actually few ways we force AI to follow rules. Here are the two you’ll see everywhere ,

RAG(Retrieval-Augmented Generation) — It’s like an open book test. For example those customer support bots on banking websites. when you ask “how do i cancel my card?”, the AI doesn’t just guess. It frantically searches the bank’s 500-page PDF policy, finds page 42, and summarizes it for you.
And if the PDF is outdated or the search finds the wrong page, the AI will confidently lie to your face :|

Prompt Chaining — It’s like an assembly line. Instead of building cars, it’s cranking out content. For example, those content repurposing automations everyone flexes on Twitter, You upload a YouTube video, and it kicks off a relay race of prompts that hand off the work until the content is finished.
Step 1: Transcribe the video.
Step 2: Summarize the transcript.
Step 3: Turn the summary into a tweet.
And here’s the catch: the whole thing is completely blind. You could upload a 10-minute video of pure wind noise, and the workflow wouldn’t even flinch. It would dutifully transcribe “[wind blowing]”, then proudly spit out a tweet like, “Top 5 insights from Wind: Whoosh. #Nature #Hustle” .

See? Useful, but dumb. They just follow orders.

back to the topic. So, how do we get an AI that actually thinks for itself?

This is where AI agents enter the picture.
So the difference between a workflow and an Agent is Reasoning. In workflows, we defines the steps then AI follows them. But an AI Agent will figure out the steps when we define the goal for them.

if the human has to hold the AI’s hand and tell it exactly which tool to use, it’s not an agent. An agent is like a really good personal assistant. If you tell an assistant, “Book me a dentist appointment”, you don’t have to say, “First, unlock your phone. then, open my contacts. then, press call”.
They just figure it out.

if you ask a standard workflow to “Plan a trip to Tokyo” it will just blindly list the top 5 hotels and flights. It doesn’t even care if a typhoon is coming.

But an AI Agent thinks like a paranoid traveler. for example,

Thought : “okay, let’s check flights to Tokyo for next week”
Observation : “Yikes, flights are $3000. That’s over budget”
Decision : “I’m not booking that. Let’s check flights to Osaka instead and take the bullet train. It’s cheaper”
Thought : “Now let’s check the weather”
Observation : “It’s going to rain all weekend”
Decision : “Cancel the park visit. I’ll book tickets to the indoor aquarium instead”

That’s the magic. It just didn’t follow a list. It hit a wall — high prices, bad weather. Then backed up, and found a new path then it reasoned its way to a solution.

What is an Agent actually made of ?

I learned that an agent needs three specific things to work. To explain this, I like to think of it like the genius doctor I met at the hospital recently. I’ll use him as an analogy.

1. The Brain (The Model) — This is Dr. Genius, he has memorized every medical textbook since 1850. He can diagnose a rare tropical disease just by looking at your eyebrows. But… he has zero street smarts. If you ask him where the cafeteria is, he panics. He knows medicine, he doesn’t know context.
   So, this is what we called the LLM that we discussed earlier. It has all the knowledge, but it needs direction.
2. The Tools — This is the medical equipment. The stethoscope, the X-Ray machine, the prescription pad etc. Without these, Dr. Genius is just a guy in a white coat staring at you awkwardly. He needs tools to actually examine you and do something about it.
   These are the capabilities we give the AI such as web search, Calculator, API access etc.
3. The Brainstem (The Prompt) — This is the Head Nurse. She is the one standing behind Dr. Genius whispering, “Doctor, check his blood pressure, don’t just guess” or “Doctor, stop staring at the patient..”. She gives the genius a specific mission and keeps him on track.
   It’s the System Prompt. It tells the model what to do, what rules to follow, and how to behave.

Based on a true story :|

How do Agents “Think”?

Okay, so we have the parts. But how do we actually make them work? Just having a brain and tools isn’t enough the agent needs a strategy. I found out that developers use specific design patterns to give agents their reasoning powers.

Planning: Instead of jumping into a task , the Agent pauses. It creates a To-Do List and breaks your big goal into small, manageable steps before it even types a single word.

Reflection (Self-Correction): You basically tell the AI to critique its own work. It’s like when you write a risky text to your crush and then read it over five times to make sure you don’t sound crazy (don’t judge me). If the agent sees a mistake, it fixes it itself before you ever see it.

Reasoning Loops (ReAct): Remember the Tokyo trip example? That’s called a ReAct pattern (Reason + Act).

Why Stop at one?

Now, here is where it gets really crazy. If one agent is powerful, what happens when you put them in a room together?

Multi-Agent Systems

You might wonder, “Why not just have one super-smart AI do everything?” Well, imagine building a house. Would you hire one guy to be the Architect, the Plumber, the Electrician, and the Painter?
Well, he’d be exhausted, and your plumbing would probably leak electricity.

Just like in a company, AI works better when it specializes.
Agent A is the Researcher.
Agent B is the Writer.
Agent C is the Editor.
They talk to each other to get the job done. This can happen in different ways.
Sequential: Like a relay race (Agent A passes to Agent B).
Hierarchical: A boss agent assigns tasks to worker agents.
Asynchronous: Everyone works at the same time and meets up at the end.

In short, we are moving from an era of search to an era of execution.

This is the foundational logic of AI Agents. We’ve stopped building scripts that the AI must follow and started building brains that we can finally trust to lead. It’s a shift from ‘How do I do this?’ to ‘Here is the goal, go make it happen.’

Honest opinion?… It’s a little scary to think about how fast this is moving, but honestly? It’s just really cool.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI