Agentic AI 103: Building Multi-Agent Teams

Introduction

articles right here in TDS, we explored the basics of Agentic AI. I’ve been sharing with you some ideas that can enable you navigate via this sea of content material we now have been seeing on the market.

Within the final two articles, we explored issues like:

create your first agent
What are instruments and implement them in your agent
Reminiscence and reasoning
Guardrails
Agent analysis and monitoring

Good! If you wish to examine it, I recommend you have a look at the articles [1] and [2] from the References part.

Agentic AI is without doubt one of the hottest topics in the mean time, and there are a number of frameworks you’ll be able to select from. Luckily, one factor that I’ve seen from my expertise studying about brokers is that these elementary ideas may be carried over from one to a different.

For instance, the category Agent from one framework turns into chat in one other, and even one thing else, however often with related arguments and the exact same goal of connecting with a Massive Language Mannequin (LLM).

So let’s take one other step in our studying journey.

On this submit, we are going to learn to create multi-agent groups, opening alternatives for us to let AI carry out extra complicated duties for us.

For the sake of consistency, I’ll proceed to make use of Agno as our framework.

Let’s do that.

Multi-Agent Groups

A multi-agent staff is nothing greater than what the phrase means: a staff with a couple of agent.

However why do we want that?

Effectively, I created this easy rule of thumb for myself that, if an agent wants to make use of greater than 2 or 3 instruments, it’s time to create a staff. The explanation for that is that two specialists working collectively will do a lot better than a generalist.

Once you attempt to create the “swiss-knife agent”, the likelihood of seeing issues going backwards is excessive. The agent will develop into too overwhelmed with completely different directions and the amount of instruments to cope with, so it finally ends up throwing an error or returning a poor end result.

Then again, while you create brokers with a single function, they are going to want only one device to unravel that drawback, subsequently rising efficiency and bettering the end result.

To coordinate this staff of specialists, we are going to use the category Crew from Agno, which is ready to assign duties to the right agent.

Let’s transfer on and perceive what we are going to construct subsequent.

Undertaking

Our challenge can be targeted on the social media content material technology business. We’ll construct a staff of brokers that generates an Instagram submit and suggests a picture primarily based on the subject offered by the person.

The person sends a immediate for a submit.
The coordinator sends the duty to the Author
- It goes to the web and searches for that subject.
The Author returns textual content for the social media submit.
As soon as the coordinator has the primary end result, it routes that textual content to the Illustrator agent, so it could actually create a immediate for a picture for the submit.

Workflow of the Crew of brokers. Picture by the creator.

Discover how we’re separating the duties very effectively, so every agent can focus solely on their job. The coordinator will guarantee that every agent does their work, and they’re going to collaborate for ultimate end result.

To make our staff much more performant, I’ll prohibit the topic for the posts to be created about Wine & Positive Meals. This manner, we slender down much more the scope of data wanted from our agent, and we will make its function clearer and extra targeted.

Let’s code that now.

Code

First, set up the mandatory libraries.

pip set up agno duckduckgo-search google-genai

Create a file for atmosphere variables .env and add the wanted API Keys for Gemini and any search mechanism you’re utilizing, if wanted. DuckDuckGo doesn’t require one.

GEMINI_API_KEY="your api key"
SEARCH_TOOL_API_KEY="api key"

Import the libraries.

# Imports
import os
from textwrap import dedent
from agno.agent import Agent
from agno.fashions.google import Gemini
from agno.staff import Crew
from agno.instruments.duckduckgo import DuckDuckGoTools
from agno.instruments.file import FileTools
from pathlib import Path

Creating the Brokers

Subsequent, we are going to create the primary agent. It’s a sommelier and specialist in connoisseur meals.

It wants a identify for simpler identification by the staff.
The function telling it what its specialty is.
A description to inform the agent behave.
The instruments that it could actually use to carry out the duty.
add_name_to_instructions is to ship together with the response the identify of the agent who labored on that activity.
We describe the expected_output.
The mannequin is the mind of the agent.
exponential_backoff and delay_between_retries are to keep away from too many requests to LLMs (error 429).

# Create particular person specialised brokers
author = Agent(
    identify="Author",
    function=dedent("""
                You might be an skilled digital marketer who focuses on Instagram posts.
                You understand how to jot down an enticing, Search engine optimization-friendly submit.
                You realize all about wine, cheese, and connoisseur meals present in grocery shops.
                You might be additionally a wine sommelier who is aware of  make suggestions.
                
                """),
    description=dedent("""
                Write clear, participating content material utilizing a impartial to enjoyable and conversational tone.
                Write an Instagram caption concerning the requested {subject}.
                Write a brief name to motion on the finish of the message.
                Add 5 hashtags to the caption.
                When you encounter a personality encoding error, take away the character earlier than sending your response to the Coordinator.
                        
                        """),
    instruments=[DuckDuckGoTools()],
    add_name_to_instructions=True,
    expected_output=dedent("Caption for Instagram concerning the {subject}."),
    mannequin=Gemini(id="gemini-2.0-flash-lite", api_key=os.environ.get("GEMINI_API_KEY")),
    exponential_backoff=True,
    delay_between_retries=2
)

Now, allow us to create the Illustrator agent. The arguments are the identical.

# Illustrator Agent
illustrator = Agent(
    identify="Illustrator",
    function="You might be an illustrator who focuses on photos of wines, cheeses, and nice meals present in grocery shops.",
    description=dedent("""
                Based mostly on the caption created by Marketer, create a immediate to generate an enticing photograph concerning the requested {subject}.
                When you encounter a personality encoding error, take away the character earlier than sending your response to the Coordinator.
                
                """),
    expected_output= "Immediate to generate an image.",
    add_name_to_instructions=True,
    mannequin=Gemini(id="gemini-2.0-flash", api_key=os.environ.get("GEMINI_API_KEY")),
    exponential_backoff=True,
    delay_between_retries=2
)

Creating the Crew

To make these two specialised brokers work collectively, we have to use the category Agent. We give it a reputation and use the argument to find out the kind of interplay that the staff may have. Agno makes accessible the modes coordinate, route or collaborate.

Additionally, don’t overlook to make use of share_member_interactions=True to guarantee that the responses will stream easily among the many brokers. It’s also possible to use enable_agentic_context, that permits staff context to be shared with staff members.

The argument monitoring is sweet if you wish to use Agno’s built-in monitor dashboard, accessible at https://app.agno.com/

# Create a staff with these brokers
writing_team = Crew(
    identify="Instagram Crew",
    mode="coordinate",
    members=[writer, illustrator],
    directions=dedent("""
                        You're a staff of content material writers working collectively to create participating Instagram posts.
                        First, you ask the 'Author' to create a caption for the requested {subject}.
                        Subsequent, you ask the 'Illustrator' to create a immediate to generate an enticing illustration for the requested {subject}.
                        Don't use emojis within the caption.
                        When you encounter a personality encoding error, take away the character earlier than saving the file.
                        Use the next template to generate the output:
                        - Put up
                        - Immediate to generate an illustration
                        
                        """),
    mannequin=Gemini(id="gemini-2.0-flash", api_key=os.environ.get("GEMINI_API_KEY")),
    instruments=[FileTools(base_dir=Path("./output"))],
    expected_output="A textual content named 'submit.txt' with the content material of the Instagram submit and the immediate to generate an image.",
    share_member_interactions=True,
    markdown=True,
    monitoring=True
)

Let’s run it.

# Immediate
immediate = "Write a submit about: Glowing Water and sugestion of meals to accompany."

# Run the staff with a activity
writing_team.print_response(immediate)

That is the response.

Picture of the Crew’s response. Picture by the creator.

That is how the textual content file seems like.

- Put up
Elevate your refreshment sport with the effervescence of glowing water! 
Overlook the sugary sodas, and embrace the crisp, clear style of bubbles. 
Glowing water is the last word palate cleanser and a flexible companion for 
your culinary adventures.

Pair your favourite glowing water with connoisseur delights out of your native
grocery retailer.
Strive these pleasant duos:

*   **For the Traditional:** Glowing water with a squeeze of lime, served with 
creamy brie and crusty bread.
*   **For the Adventurous:** Glowing water with a splash of cranberry, 
alongside a pointy cheddar and artisan crackers.
*   **For the Wine Lover:** Glowing water with a touch of elderflower, 
paired with prosciutto and melon.

Glowing water is not only a drink; it is an expertise. 
It is the right option to get pleasure from these particular moments.

What are your favourite glowing water pairings?

#SparklingWater #FoodPairing #GourmetGrocery #CheeseAndWine #HealthyDrinks

- Immediate to generate a picture
A vibrant, eye-level shot inside a connoisseur grocery retailer, showcasing a range
of glowing water bottles with numerous flavors. Organize pairings round 
the bottles, together with a wedge of creamy brie with crusty bread, sharp cheddar 
with artisan crackers, and prosciutto with melon. The lighting must be shiny 
and welcoming, highlighting the textures and colours of the meals and drinks.

After we now have this textual content file, we will go to no matter LLM we like higher to create photos, and simply copy and paste the Immediate to generate a picture.

And here’s a mockup of how the submit can be.

Mockup of the submit generated by the Multi-agent staff. Picture by the creator.

Fairly good, I’d say. What do you assume?

Earlier than You Go

On this submit, we took one other step in studying about Agentic AI. This subject is sizzling, and there are various frameworks accessible out there. I simply stopped making an attempt to study all of them and selected one to start out truly constructing one thing.

Right here, we had been in a position to semi-automate the creation of social media posts. Now, all we now have to do is select a subject, modify the immediate, and run the Crew. After that, it’s all about going to social media and creating the submit.

Definitely, there may be extra automation that may be performed on this stream, however it’s out of scope right here.

Concerning constructing brokers, I like to recommend that you just take the better frameworks to start out, and as you want extra customization, you’ll be able to transfer on to LangGraph, for instance, which permits you that.