TopolearnAgent definitions are prompts in disguise; good ones are the biggest lever on output quality.
An Agent is one AI worker. CrewAI turns three plain-text fields into the agent's instructions:
role: the job title, for example "Senior market researcher". It appears in prompts and logs.goal: what the agent is trying to achieve, in one or two sentences.backstory: background that sets tone and expertise, such as "You have 10 years of experience analysing consumer apps and you always cite sources."
Together these act like a detailed system prompt. Specific, concrete wording works far better than vague praise like "You are the best researcher".
Other useful settings:
llm: which model this agent uses (a later lesson). If you leave it out, CrewAI uses theMODELenvironment variable or its default OpenAI model.tools: functions the agent may call, like web search.verbose=True: print the agent's reasoning and tool use, very helpful while learning.allow_delegation: lets the agent hand work to other agents in the crew. It isFalseby default.max_iter: the most think-act cycles the agent may use on one task (default 25), a safety limit against endless loops.max_rpm: a cap on model requests per minute, to respect provider rate limits.
Text fields can contain {placeholders} such as {topic}. They are filled in when the crew starts, so one agent definition can serve many inputs.
An agent does nothing on its own: it needs a task to work on.
Example
from crewai import Agent
researcher = Agent(
role="Senior {industry} researcher",
goal="Find recent, verifiable facts about {topic} and cite every source",
backstory=(
"You have ten years of experience in {industry} analysis. "
"You prefer primary sources and say clearly when something is uncertain."
),
verbose=True, # show reasoning and tool calls in the terminal
allow_delegation=False, # default: do not hand work to other agents
max_iter=15, # stop after 15 think-act cycles on a task
)
print(researcher.role) # Senior {industry} researcher (filled in at kickoff)Remember
- role, goal and backstory become the agent's instructions
- Be concrete: expertise, standards and constraints beat flattery
max_iter,max_rpmandallow_delegationcontrol cost and behaviour
Watch out for
- Writing a long, generic backstory; it costs tokens on every call and rarely helps more than two specific sentences.