The Team: Orchestrating Crews and Processes

The Team: Orchestrating Crews and Processes

We have Agents (the workers) and Tasks (the work). The Crew is the entity that brings them together and manages the execution flow.

Assembling the Crew

A Crew is initialized with a list of agents and a list of tasks. It acts as the container for the entire operation.

from crewai import Crew, Process

my_crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],
    verbose=True
)

Processes: The Workflow Styles

How should the tasks be executed? CrewAI supports different Process types.

1. Sequential (Default)

This is the “Waterfall” method. Tasks are executed in the order they are listed in the tasks list. The output of the first task is automatically passed as context to the second task, and so on. This is ideal for linear workflows.

[Task A] -> Output -> [Task B] -> Output -> [Task C]

2. Hierarchical

This simulates a managerial structure. You define a list of tasks, but instead of assigning them manually, a “Manager” agent (which CrewAI creates for you) analyzes the tasks and delegates them to the available agents based on their roles. To use this, you must specify a manager_llm.

my_crew = Crew(
    agents=[researcher, writer],
    tasks=[task1, task2],
    process=Process.hierarchical,
    manager_llm=gpt4_llm
)

Kickoff: Starting the Engine

To start the process, you call the kickoff() method. This is where you pass inputs for any variables defined in your tasks/agents.

inputs = {'topic': 'Quantum Computing'}
result = my_crew.kickoff(inputs=inputs)

print(result)

The result is a CrewOutput object, containing the final answer, token usage statistics, and individual task outputs.

Memory and Caching

By setting memory=True in the Crew config, you enable agents to remember past interactions (Short-term memory) and recall information across different runs (Long-term memory). We will cover this in depth in a later chapter, but know that the Crew class handles this persistence layer.