Your First AI Agent Team

I'm sure you've already seen those AI agent diagrams flooding your LinkedIn feed.

You know the ones: "Check out this AI agent that's making $50K per month! Drop 'AGENT' in the comments for my free template..."

The more complex the agent, the more people click as they say, "oh this is hard to replicate!"

Last week, I spent some time with one of these "advanced" AI agent workflows.

It had 47 nodes.

Took me a solid 10 minutes to understand what it even did.

And after all that complexity?

It was slower than my brain before coffee.

We've made AI too complicated. Some people think more complexity means more value — that's 100% not true.

I've actually found the opposite to be true: the simplest solutions tend to be the most powerful. When something's overly complex, it's usually just not well-designed yet.

Today, we're flipping the script. I'm going to show you how to build simple agent systems, but genuinely useful in your daily work. You’ll build your first multi-agent system.

And don't worry — you won't need a computer science degree to follow along.

What Makes an AI Agent Effective?

An Agent in a nutshell is an AI system that:

• Makes decisions
• Uses tools and APIs
• Has memory & adapts on the fly

Unlike a workflow that executes the exact same sequence of tasks every time.

What matters:

• A clearly defined job
• The right tools
• A clear input
• A simple, reliable workflow
• Guardrails
• Output format
• Error handling

Think of it like hiring an employee.

Would you give them a vague job title, zero instructions, and expect greatness?

There's another crucial component we need to add: trust. It's the real challenge here.

Trust boils down to three simple questions:

• Can I stake important decisions on this AI? (Like with my finances, health, or safety?)
• Can I understand what it's doing and why it's doing it?
• Does this system actually align with what matters to me?

I won't dive deeper into trust today, but just know—it's currently the biggest bottleneck for AI agents becoming truly useful.

A Framework That Works: Build Your AI Work Team

This is not ok ↓

Intractable, hard to maintain, hard to improve, and in 90% of cases very slow and inefficient.

I'll show you how to create an efficient Agent architecture using multi-agent systems.

Let’s break it down.

We’re going to use a building block approach —

Where each AI agent is a specialist in one thing, and a manager coordinates it all:

• One agent does research
• Another handles data visualization
• A manager agent assigns the tasks

All built on n8n — a no-code platform. If you've never used n8n before, this is your quick start tutorial.

(n8n is not the only platform to connect agents, but the same principles apply to any other platform)

Step 1: Map Your Workflow

List 2-3 key, repetitive tasks you do regularly.

Examples:

• Market research
• Competitor analysis
• Data visualization
• Drafting emails
• Report generation

You want tasks that are:

• Specific
• Repeatable
• Annoying to do manually

Step 2: Build Specialized Agents (The Workers)

Let’s build our first two agents.

One simple hack that saves me a lot of time is to have an "Agent Designer Assistant." This Agent Designer will help you craft good prompts for whatever agents you want to create, and you don't have to spend dozens of minutes creating a "good" prompt for your agents.

You simply describe what you need in plain English, and the Agent Designer crafts a well-structured prompt that gets results. No need to spend ages perfecting your prompts.

You can use this prompt:

ENERATE AN AGENT PROMPT WITH:

### **ROLE & RESPONSIBILITY**
- Primary function and scope  
- Key deliverables expected  
- Authority/decision-making boundaries  

---

### **AVAILABLE TOOLS**
- Tool name: Specific use cases  
- When to use each tool  
- Tool limitations/constraints  

---

### **WORKFLOW PROCESS**
1. Input analysis and validation  
2. Step-by-step execution sequence  
3. Quality checks at each stage  
4. Output formatting and delivery  

---

### **IMPORTANT GUARDRAILS**
- What the agent must never do  
- Error handling protocols  
- Escalation triggers for human intervention  

---

### **OUTPUT SPECIFICATIONS**
- Format requirements  
- Delivery method  
- Follow-up actions needed  

---

STYLE: Professional, actionable, specific. Include concrete examples where helpful._

🤖 Agent 1: The Research Assistant

What it does:

• Uses Perplexity to do competitive research
• Summarizes the insights
• Writes a report in Google Docs
• Returns the link

(I can copy paste this simple description to my "Agent Designer" on Claude to create a clear prompt.)

Here’s the next steps:

1. Design the agent prompt
2. Use an AI agent designer (mine is on Claude, cf. above).
3. Make sure that the “AVAILABLE TOOLS” section is clear and well crafted.
4. Set up tools in n8n

This is where people can get lost quickly so let me give you a little more detail.

First, go to n8n and create an account — choose a free trial to get started.

→ Second, you need to create your workflow:

→ Then add your AI Agent:

Now you can clearly see the structure and requirements of an AI agent:

• A chat model
• Memory
• Tools

→ For the AI model, I'll pick Claude 4 — make sure to activate "reasoning". For credentials, you'll just need to create and add an API key from whatever AI platform you're using (super easy, you need a credit card, ask ChatGPT or any AI assistant to guide you). The price is peanuts.

→ For the memory, I'll just pick a simple memory and set it to 5 (which means I'm only focusing on the last five messages)

→ For the tools, I'll go to my System Prompt generated by Claude above and just add all the tools mentioned in the prompt. In this case, I need the Perplexity Tool. You'll need an API key from Perplexity—for the other settings, you can keep them at default.

→ You can also add any other relevant tool, like access to a Google Drive (again, you'll just need an API key from your Google platform). For this specific tool, I want to read from a given document, so I'll choose the "Get" operation and simply provide the URL to my doc ↓

→ Now I can add another node to create a Google Drive document whenever the research is completed. I can either update an existing document or create a new one for each research. It's pretty straightforward on n8n ↓

→ Finally I'll add my system prompt — generated by my Claude assistant — to my agent.

Keep in mind that the system prompt is a crucial component. The agent will decide what to do with the tools and when based on that description. Make sure your system prompt is rock solid (review and refine Claude's outputs).

(I'm using the SaaS free version for demonstration purposes and I haven't added my credentials. Once your API keys are active, all your tools will be up and running and you can simply test your agent)

Example Task

“Research Figma’s business model, products, and market position, and how it relates to Miro.”

⏱ Time to build: under 15 minutes

🧾 Output: Clean, organized Google Doc

🔍 Insights + strategy recommendations

Pro tip: Adjust your system prompt to include formatting preferences (word count, structure, etc.)

Now, try building the agent below by following the exact same steps outlined above.

📊 Agent 2: The Visualization Assistant

What it does:

• Takes a Google Doc report
• Turns it into a dashboard
• Sends it to you via email

Again, we start with a prompt.

Ask your “Agent Designer Assistant”:

“Create an agent that turns a Google Doc into a visual dashboard and emails it to me.”

Then:

1. Use the same model (Claude Sonnet 4)
2. Increase the token limit (set to 8,000)
3. Add Google Doc (Read) + Gmail (Send) nodes
4. Let the AI model decide document ID, email content, subject

Test everything.

Make it beautiful.

🎯 Result: An email in your inbox with a summary + visuals

The Final Boss: The Manager Agent

This is where it all comes together.

A manager agent = your AI project lead

It delegates, coordinates tasks, and ensures your tiny AI team works together seamlessly.

Here’s what it does:

• Calls the research agent
• Waits for the output
• Passes the output to the visualization agent
• Notifies you when everything’s done

Let me quickly show you how to set this up.

→ Setup

1. Create a system prompt for the manager (again using your Agent Prompt Designer)

Example: “Your job is to delegate tasks to the research and visualization agents and deliver results.”

2. Modify the two existing agents:

- Remove memory

- Remove chat input trigger

- Replace with "Execute by another workflow"

- Add input field (query)

- Link their prompts to this query using {{$json.query}}

3. In the manager agent:

- Create the Manager Agent (as a new agent, following the same steps as before)

- Add “Call Workflow” tools for both sub-agents

- Define the query field for each

- Paste in the system prompt

🎯 Now you can send a single prompt to the manager agent like:

“Research FigJam and send me a dashboard with strategic insights.”

The rest is automated.

IX. Your Next Step

You already saw the power of a three-agent team.

But you can go further.

Think:

• Calendar assistants
• Proposal writers
• Sentiment analyzers
• Outreach campaign bots
• SEO monitors

The building blocks are the same. Only the instructions change.

But please don't build 47 connections workflows. That will make your life harder, not easier.

Final Thought

You don’t need to wait for “agentic AI” to arrive. It’s already here. You just have to build it.

And remember:

Complex ≠ Smart
Simple ≠ Basic
Effective = Clear + Focused + Modular

I hope this simple guide has convinced you that building your first (well-designed) agent takes just one hour with n8n and this short letter.

For those of you who are already familiar with building n8n agents (and may have found this letter boring -_-), take a moment to reflect on your approach. Are you building overly complex systems that nobody else can understand? Difficult to maintain? Or are you creating simple, modular agents connected through manager workflows that anyone can follow and modify?

Wishing you a great weekend.

Until the next one,

— Charafeddine