SelfManaged Crew

AI crews allow you to use entire teams of AI agents to perform complex tasks. It may initially seem confusing, but the crew approach simply copies how real teams work. In any given team, you’ll have individuals with unique roles and skills, working together to reach a common goal.

Let’s say you want to create and publish a long-form blog post. The work usually starts with an SEO specialist researching keywords and outlining the content. They’ll create an SEO brief, which will be passed on to the content writer. Once the writer is done, their colleague will proofread and edit the article to ensure quality. What about the featured images or infographics? A designer will help with that. 

You already have at least three or four people working on creating this piece of content. They share a common goal, but each specializes in something else and performs a different subtask.  Let’s see how you can copy this team as a group of AI agents.

Curious about the Flow we’re analyzing in this guide? It’s the Advanced Blog Generator and you can easily find it in your Flow library. 

The Role of AI Agents in Crews

The SelfManaged Crew component is only a structural component bringing groups of agents together. Because of this, the first step to successfully using AI Crews is understanding and setting up individual agents, including the manager agent. 

AI Agents are computer programs that can perform tasks and solve problems independently. They process information and take action based on their programming, knowledge, and goals. 

Agents aren’t just generative AI. Given the right tools, they can perform real tasks like sending emails, creating documents, and more. Instead of pre-defining rigid triggers for this behavior, Agents can decide independently. 

In practice, you no longer need to give detailed prompts for a rigidly controlled generative behavior. All you need to do is provide an agent with their role, personality, and goal, ensuring they know who they are and what motivates them. 

Learn more about AI agents and how to use the AI Agent component 

How Are Crews Better Than Single Agents?

If there’s an issue in your team’s processes, you can quickly pinpoint the issue and work with a competent team member to find a solution. Now imagine it’s just you working on the entire task, and the issue arises in your own mind. That’s much harder to notice and pinpoint. The same happens when comparing a single agent with a crew of agents. 

When prompting a single agent, you give it a complex task with little to no control over how individual subtasks are performed. When doing complex tasks, this can lead to bottlenecks and low output quality. 

With a crew, you can split the main task into specific subtasks, assigning each to a unique AI team member. The result is a much more professional and detailed output. It also means easier debugging, and lastly, coordinating specialized agents allows you to handle much more complex tasks. 

The Difference Between Self-Managed and Sequential Crews

You might have noticed there are two different Crew components in your dashboard. The difference between these types of crews is in the order of tasks and the level of control you get.

Let’s go back to our marketing team example. The first agent in line would be the SEO specialist. Once the topic is researched, it forwards the information to the content writer. Below, you can see how the SEO Agent’s task is connected to the Writing task of the Content Writer Agent:

Let’s talk about a Sequential Crew first. With a Sequential Crew, the tasks are performed one after another in the exact order you specify in the Flow. Once a task is done, it’s permanent, and the Flow moves on to the next agent. That is great for straightforward processes or processes that require less computational power.

Let’s focus on a real-life content writer. They will first do research and move on to writing, but as the article unfolds, they may realize more research is needed. Understandably, they will go back and forth between research and writing tasks before finally moving on to the next step. The sequential crew won’t do this. Once a task is done, it’s just done. That’s where Self-Managed crews come in.

With a Self-Managed Crew, the manager AI agent decides the order of tasks and how many iterations are needed. When making decisions, the AI tries to copy traditional organizational hierarchies closely. This opens up the possibility of repeating tasks and creating multiple iterations of the final output. 

Thanks to the manager LLM that delegates tasks and oversees their execution, the SelfManaged Crew can work with a single complex task. The manager LLM can seamlessly split the task and assign subtasks to the correct agents. This is especially great when you know what needs to be done, but you’re not sure about the exact process and subtasks. 

How To Use SelfManaged Crews

The SelfManaged Crew is a structure component that brings agents and task components together in a group. To use a SelfManaged Crew, we need to first define the manager agent, the team members, and their tasks. Only then can we make them a team.

Setting up SelfManaged Crews consists of four steps: 

1. Setting up individual AI Agents
2. Giving agents tasks
3. Setting up the manager agent
4. Making the agents a SelfManaged Crew 

Setting Up Individual AI Agents 

Each member of a real team has a role, goals, and a unique backstory that includes their past experiences, personality, and specific style. So does each AI Agent. 

For example, let’s focus on the content writer team member: 

• The Role is your agent’s job title. In this example, being a content writer is the role.  
• The Goal is what the agent does and what their ideal outcome is. The expected outcome for the content writer is a well-written article that adheres to the theme and SEO brief. 
• The Backstory represents who the agent is. Whether you like it or not, you always bring your personality, way of thinking, vocabulary, and past experiences to anything you do. This is even more visible in creative work, such as content writing. 

Repeat this process for all the agents you want to use in your team. 

Learn more about AI agents and how to use the AI Agent component 

Giving Agents Tasks

Continuing with our blog creation example, we now know who our agent is. The next step is to let the agent know their task and introduce them to the team. 

What Are The Task Components? 

In Crews, each Agent is assigned a task to perform. Like in a real team, each member can carry out various project-specific tasks. The task components allow you to specify and assign these tasks. 

You’ll notice that, like with the Crew component, there are two possible task components— sequential and SelfManaged. Since these are two opposite approaches to managing agents, mixing them would make no sense. That’s why we’ll also use SelfManaged Tasks when using a SelfManaged Crew: 

If you have a task in mind but are unsure how to split it into smaller subtasks, simply write it all into a single task. The manager LLM is there to assign tasks and oversee the process, ensuring each agent knows what to do and when. It can split the main task and assign the parts to the correct agent.

In addition to the task, each agent in a Crew can also get appropriate tools, making their job easier and more accurate. In our example, the researcher uses the GoogleSearch and URL Retriever tools to control the research options. 

Next, set up the tasks. Each SelfManaged Task must either have a description, the expected output, or both:  

The task description for the content writer agent might go a little something like this: 

“Given the SEO content brief, write a blog post in no more than 1500 words. 

Never start paragraphs with vague statements such as "In the fast-changing field of...". Always go directly to the main information the paragraph should deliver. “

Let’s take a closer look at this task description:

“Given the content brief” – The agent knows what to do with the previous output.

“Write a blog post of up to 1500 words” = The output we expect from the agent.

“Never start…..” = Giving additional custom instructions to tweak the output. These instructions can be any pointers on language, vocabulary, structure, or anything else that will help the agent create what you need.  

The expected output field is optional and works great when you need a structured output or make sure something is included in the output. For example, our SEO researcher agent’s task is to create: 

A brief in this form:

SEO friendly Title:

SEO friendly Meta description:

SEO friendly Outline

Ensuring it doesn’t forget to start the output with a title and meta description. 

The final step is to connect all tasks into the SelfManaged Crew component. From there, the manager LLM will take over, ordering the tasks as necessary to achieve the expected output. 

Remember that sequential crews work with sequential tasks only, and SelfManaged crews work with SelfManaged tasks only.

Setting Up The Manager Agent 

Thinking back to real-life teams, how do they deal with ensuring smooth communication and the execution of complex tasks? Easy, they introduce a manager, whose main task is to oversee completion and give out tasks while facilitating communication between team members. That’s what the manager agent does in your AI agent team. 

Unlike Sequential Crews, which work in a fixed sequence you set, SelfManaged Crews can work dynamically and iterate tasks according to new information. This requires a manager to define and oversee the process, bsetting up the manager LLM isn’t much different from setting up all the other agents. 

Set the manager agent up as you would any other. Give it a backstory, goal, and role. For example, our manager LL is a content team lead with great people skills, and their task is to oversee the content creation process.

Then connect it to the SelfManaged Crew component as the Manger Agent :  

All agents contain the newest version of ChatGPT by default. Only connect an LLM component if you wish to change to a different AI model. 

Making The Agents A Crew 

Let’s go back to our Flow. It features three team member agents and a task for each. Overseeing how they work is the manager agent. The last step of creating a crew is letting the agents know they’re a team. This is where the SelfManaged Crew component comes into play. 

The SelfManaged Component 

The SelfManaged Crew component represents a group of agents whose work is automatically managed by a manager LLM. The team manages itself, allowing for dynamic work and the ability to create several iterations. It’s essentially a way to tell agents they’re a team with a common goal. 

There may be more than one independent team within your Flow, meaning more than one crew component distinguishing these teams from each other. In our example, we only use one crew, but we still need to bring the agents together in a crew: 

Let’s look at this component’s settings:

Input Handles

Agents: Connect all the agents that belong to this crew.

Manager Agent: Connect the agent whose task it is to be the manager (don’t forget to properly set up the manager agent)

Manager LLM: All agents include the newest ChatGPT version by default. If you want to pick any other of the dozens of LLMs, simply drag it in and connect it to the LLM handle.  

Tasks: Connect all the tasks that the crew will perform.

Output Handles

Output: The component outputs text, ready to go to chat output or for further processing.

Settings

Show Progress of Agent: Check if you want to see a detailed log of how the agents think and how they approach tasks. This is a great tool for debugging, as it lets you pinpoint the weak spots of your workflow. 

That’s it. Just send it to output, and now you have a team of agents working in an exact order. Our Flow includes three agents: an SEO specialist, a content writer, and an editor. 

The Flow used in this was the Advanced Blog Generator you can find in your Flow library.

Instead of vague output laden with telltale AI phrases, this Flow’s output will be well researched, inspired by top Google results, written according to a clear brief, and edited to avoid sounding like generic AI.  Plus, using a group of agents instead of a single agent minimizes bottlenecks. It ensures that any issues can be diagnosed and resolved promptly by simply tweaking one of the agents.