Your first crew: agents and tasks

A crew is a team that runs a list of tasks with a list of agents and returns one result; kickoff starts it.

Last updated: 28 Sep, 2026 · CrewAI 1.15

ShopLLM from the custom-LLM lesson answers messages. An agent wraps it with a job, a task says what the job is, and a crew runs them together.

ShopLLM answers messages. An agent wraps it with a job description, in three parts CrewAI requires: a role, a goal and a backstory.

import os
import re

from crewai import BaseLLM

os.environ["CREWAI_DISABLE_TELEMETRY"] = "true"
os.environ["CREWAI_TRACING_ENABLED"] = "false"
os.environ["CREWAI_DISABLE_VERSION_CHECK"] = "true"


class ShopLLM(BaseLLM):
    def call(self, messages, tools=None, **kwargs):
        if isinstance(messages, str):
            messages = [{"role": "user", "content": messages}]
        text = messages[-1]["content"]
        orders = re.findall(r"\b[A-Z]\d+\b", text)
        if orders:
            return f"I have no way to look up {orders[0]} yet."
        return "Hello. Which order is this about?"

An agent

from crewai import Agent
from shop_llm import ShopLLM

agent = Agent(
    role="Support agent",
    goal="Answer customers of a small online shop",
    backstory="You have worked the shop's support desk for years.",
    llm=ShopLLM(model="shop"),
)

print(agent.role)

Support agent

Nothing ran yet. The agent is a description of a worker; it needs a task.

A task and a crew

from crewai import Crew, Task

task = Task(
    description="Answer the customer: Where is my order A17?",
    expected_output="One short, friendly sentence.",
    agent=agent,
)
crew = Crew(agents=[agent], tasks=[task])

result = crew.kickoff()
print(result.raw)

I have no way to look up A17 yet.

A task has a description of the work and an expected_output saying what finished looks like, and names the agent that does it. The crew takes lists of agents and tasks. kickoff runs the tasks and returns a CrewOutput, whose raw is the last task's answer as text.

Reading each task's output

result = crew.kickoff()

for output in result.tasks_output:
    print(output.agent, "|", output.raw)

Support agent | I have no way to look up A17 yet.

tasks_output keeps one result per task, each tagged with the role of the agent that produced it. With one task there is one entry; a crew of two agents has two.

The crew holds the lists; the task says what to do and who does it; the agent turns that into messages for the model. A run this small has four moving parts, and you wrote one of them.

An agent with no model

from crewai import Agent

agent = Agent(
    role="Support agent",
    goal="Answer customers of a small online shop",
    backstory="You have worked the shop's support desk for years.",
)
print(agent.llm.model)

gpt-4.1-mini

Leaving out llm makes CrewAI choose a hosted OpenAI model. The installed 1.15 selects gpt-4.1-mini here, printed above; the choice can change between versions, and the Agents page of the documentation still names an older default, so pass llm= yourself rather than depend on it. Creating the agent works. Running it is where the key is needed:

from crewai import Agent, Crew, Task
import shop_llm

agent = Agent(role="Support agent", goal="Answer customers", backstory="You work the desk.")
task = Task(description="Where is my order A17?", expected_output="One sentence.", agent=agent)
Crew(agents=[agent], tasks=[task]).kickoff()

[CrewAIEventsBus] Warning: Event pairing mismatch. 'llm_call_failed' closed 
'flow_started' (expected 'llm_call_started')
[CrewAIEventsBus] Warning: Event pairing mismatch. 'llm_call_failed' closed 
'flow_started' (expected 'llm_call_started')
[CrewAIEventsBus] Warning: Event pairing mismatch. 'llm_call_failed' closed 
'flow_started' (expected 'llm_call_started')
Traceback (most recent call last):
  File "main.py", line 6, in <module>
    Crew(agents=[agent], tasks=[task]).kickoff()
ValueError: OPENAI_API_KEY is required

CrewAI tried the call three times before giving up, and printed a warning line for each failed call. import shop_llm is there only for its three settings, so the failed run sends nothing. Every agent below passes llm= explicitly.

What kickoff hands back

• kickoff returns a CrewOutput; its raw is the last task's answer as text.
• tasks_output holds one entry per task, each tagged with the role of the agent that produced it.
• Leaving out llm still builds the agent; the missing key only bites at kickoff.

raw vs tasks_output

When a crew earns its place

• The work has more than one step, and each step has a clear owner.
• You want one result out, with the steps in between recorded.
• Different steps need different tools or different instructions.

Related

• Previous: A custom LLM with BaseLLM
• Next: The agent prompt: what the model is sent