Crew: Reference Guide

What is Crew?

Crew is the object you wire once and then execute: it owns the ordered list of agents, the task graph, the Process (sequential vs hierarchical), and cross-cutting services such as embedder configuration, crew-level knowledge_sources, shared memory toggles, and optional planning via AgentPlanner. kickoff() runs the full loop and returns a CrewOutput with per-task outputs and token_usage; train(), test(), and replay() support the training-and-evaluation workflow without rewriting your definitions.

Hierarchical mode is where most production surprises appear: you must supply manager_llm or manager_agent so the manager can delegate, and you should expect higher latency than sequential pipelines because the manager re-reads context between delegations. max_rpm applies globally across agents unless you also cap per-agent max_rpm — combine both when you hit provider rate limits. security_config attaches fingerprints and redaction defaults for telemetry, which matters the moment you log prompts in regulated environments.

When a Flow wraps a Crew, treat the Crew as a pure domain block: keep IO at the Flow layer and pass narrow inputs into kickoff(inputs={...}) so tasks stay deterministic. For YAML-first projects, @CrewBase still constructs a Crew under the hood — the class reference here is the imperative mirror of that layout.

When to Use

The default container for multi-agent work.

Use Cases

• • Research pipelines
• • Content factories
• • Triage systems

Key Features

• ✓ Process selection
• ✓ Shared memory & knowledge
• ✓ Manager-agent / manager-LLM
• ✓ Planning support

When NOT to Use

Notes

Import

from crewai import Crew

Key Parameters

Code Examples

from crewai import Agent, Task, Crew, Process

researcher = Agent(role='Researcher', goal='Summarize {topic}', backstory='You cite sources.')
writer = Agent(role='Writer', goal='Draft a brief', backstory='You are concise.')

t1 = Task(description='Bullet findings on {topic}', expected_output='Bullets', agent=researcher)
t2 = Task(description='Turn bullets into a 150-word brief', expected_output='Markdown', agent=writer, context=[t1])

crew = Crew(agents=[researcher, writer], tasks=[t1, t2], process=Process.sequential, verbose=True)
out = crew.kickoff(inputs={'topic': 'CrewAI planning'})
print(out.raw)

from crewai import Agent, Task, Crew, Process

worker = Agent(role='Analyst', goal='Answer', backstory='You use tools sparingly.')
reviewer = Agent(role='Reviewer', goal='Validate', backstory='You reject vague answers.')

t_do = Task(description='Analyze {case}', expected_output='Analysis', agent=worker)
t_check = Task(description='Check analysis for gaps', expected_output='Pass/fail notes', agent=reviewer, context=[t_do])

crew = Crew(
    agents=[worker, reviewer],
    tasks=[t_do, t_check],
    process=Process.hierarchical,
    manager_llm='openai/gpt-4o-mini',
    memory=False,
)
print(crew.kickoff(inputs={'case': 'Q2 churn'}).raw)

from crewai import Crew, Agent, Task, Process
from crewai.knowledge.source.string_knowledge_source import StringKnowledgeSource

policy = StringKnowledgeSource(content='Refund window is 14 days for EU customers.')
agent = Agent(role='Support', goal='Answer policy questions', backstory='You only use the knowledge source.')
task = Task(description='Does this order qualify for a refund?', expected_output='Yes/no with citation', agent=agent)

crew = Crew(
    agents=[agent],
    tasks=[task],
    process=Process.sequential,
    knowledge_sources=[policy],
    planning=True,
    planning_llm='openai/gpt-4o-mini',
)
print(crew.kickoff().raw)

Common Mistakes

Crew FAQ