CrewAI Agent Design Guide

How to design effective agents with the right role, goal, backstory, tools, and configuration.

The 80/20 Rule

Spend 80% of your effort on task design, 20% on agent design. A well-designed task elevates even a simple agent. But even the best agent cannot rescue a vague, poorly scoped task. Get the task right first (see the design-task skill), then refine the agent.

1. The Role-Goal-Backstory Framework

Every agent needs three things: who it is, what it wants, and why it's qualified.

Role — Who the Agent Is

The role defines the agent's area of expertise. Be specific, not generic.

The role directly shapes how the LLM reasons. A "Senior Data Researcher" will produce different output than a "Research Assistant" even with the same task.

Goal — What the Agent Wants

The goal is the agent's individual objective. It should be outcome-focused with quality standards.

Backstory — Why the Agent Is Qualified

The backstory establishes expertise, experience, values, and working style. It's the agent's "personality prompt."

backstory: >
  You're a seasoned researcher with 15 years of experience in AI/ML.
  You're known for your ability to find obscure but relevant papers
  and synthesize complex findings into clear, actionable insights.
  You always cite your sources and flag uncertainty explicitly.

What to include in a backstory:

• Years/depth of experience
• Specific domain knowledge
• Working style and values (e.g., "always cites sources", "prefers concise output")
• Quality standards the agent holds itself to

What NOT to include:

• Implementation details (tools, models, config)
• Task-specific instructions (those go in the task description)
• Arbitrary personality traits that don't affect output quality

2. Agent Configuration Reference

Essential Parameters

Agent(
    role="...",              # Required: agent's expertise area
    goal="...",              # Required: what the agent aims to achieve
    backstory="...",         # Required: context and personality
    llm="openai/gpt-4o",    # Optional: defaults to OPENAI_MODEL_NAME env var or "gpt-4"
    tools=[...],             # Optional: list of tool instances
)

Execution Control

Agent(
    ...,
    max_iter=25,             # Max reasoning iterations per task (default: 25)
    max_execution_time=300,  # Timeout in seconds (default: None — no limit)
    max_rpm=10,              # Rate limit: max API calls per minute (default: None)
    max_retry_limit=2,       # Retries on error (default: 2)
    verbose=True,            # Show detailed execution logs (default: False)
)

Tuning max_iter:

• Default 25 is generous — most tasks finish in 3-8 iterations
• Lower to 10-15 to fail faster when tasks are well-defined
• If agent consistently hits max_iter, the task is too vague (fix the task, not the limit)

Tool Configuration

from crewai_tools import SerperDevTool, ScrapeWebsiteTool, FileReadTool

Agent(
    ...,
    tools=[SerperDevTool(), ScrapeWebsiteTool()],  # Agent-level tools
)

Key rules:

• An agent with no tools will hallucinate data when asked to search, fetch, or read files — always provide tools for tasks that require external data
• Prefer fewer, focused tools over many tools — too many tools confuses the agent
• Tools can also be assigned at the task level for task-specific access (see design-task skill)
• Agent-level tools are available for all tasks the agent performs; task-level tools override for that specific task

LLM Selection

Agent(
    ...,
    llm="openai/gpt-4o",              # Main reasoning model
    function_calling_llm="openai/gpt-4o-mini",  # Cheaper model for tool calls only
)

Use function_calling_llm to save costs: the main llm handles reasoning while a cheaper model handles tool-calling mechanics.

Collaboration

Agent(
    ...,
    allow_delegation=False,  # Default: False — agent works alone
)

Set allow_delegation=True only when:

• The agent is part of a crew with other specialized agents
• The task genuinely benefits from the agent handing off subtasks
• You're using hierarchical process where the manager delegates

Warning: Delegation without clear task boundaries leads to infinite loops or wasted iterations.

Planning (Reasoning Before Acting)

from crewai.agents.agent_builder.base_agent import PlanningConfig

Agent(
    ...,
    planning=True,                    # Enable plan-then-execute (default: False)
    planning_config=PlanningConfig(
        max_attempts=3,               # Max planning iterations
    ),
)

Use planning for complex tasks where the agent benefits from thinking through its approach before taking action. Skip it for simple, well-defined tasks.

Code Execution

Agent(
    ...,
    allow_code_execution=True,        # Enable code execution (default: False)
    code_execution_mode="safe",       # "safe" (Docker) or "unsafe" (direct) — default: "safe"
)

• "safe" requires Docker installed and running — executes in a container
• "unsafe" runs code directly on the host — only use in controlled environments

Context Window Management

Agent(
    ...,
    respect_context_window=True,      # Auto-summarize to stay within limits (default: True)
)

When True, the agent automatically summarizes prior context if it approaches the LLM's token limit. When False, execution stops with an error on overflow.

Date Injection

Agent(
    ...,
    inject_date=True,                 # Add current date to task context (default: False)
    date_format="%Y-%m-%d",           # Date format (default: "%Y-%m-%d")
)

Enable for time-sensitive tasks (research, news analysis, scheduling).

Agent Guardrails

def validate_no_pii(result) -> tuple[bool, Any]:
    """Reject output containing PII."""
    if contains_pii(result.raw):
        return (False, "Output contains PII. Remove all personal information and try again.")
    return (True, result)

Agent(
    ...,
    guardrail=validate_no_pii,
    guardrail_max_retries=3,          # default: 3
)

Agent guardrails validate every output the agent produces. The agent retries on failure up to guardrail_max_retries.

Knowledge Sources

from crewai.knowledge.source.text_file_knowledge_source import TextFileKnowledgeSource

Agent(
    ...,
    knowledge_sources=[
        TextFileKnowledgeSource(file_paths=["company_handbook.txt"]),
    ],
    embedder={
        "provider": "openai",
        "config": {"model": "text-embedding-3-small"},
    },
)

Knowledge sources give agents access to domain-specific data via RAG. Use when agents need to reference large documents, policies, or datasets.

3. YAML Configuration (Recommended)

Define agents in agents.yaml for clean separation of config and code:

researcher:
  role: >
    {topic} Senior Data Researcher
  goal: >
    Uncover cutting-edge developments in {topic}
    with supporting evidence and source citations
  backstory: >
    You're a seasoned researcher with 15 years of experience.
    Known for finding obscure but relevant sources and
    synthesizing complex findings into clear insights.
    You always cite your sources and flag uncertainty.
  # Optional overrides (uncomment as needed):
  # llm: openai/gpt-4o
  # max_iter: 15
  # max_rpm: 10
  # allow_delegation: false
  # verbose: true

Then wire in crew.py:

@CrewBase
class MyCrew:
    agents_config = "config/agents.yaml"
    tasks_config = "config/tasks.yaml"

    @agent
    def researcher(self) -> Agent:
        return Agent(
            config=self.agents_config["researcher"],
            tools=[SerperDevTool()],
        )

Critical: The method name (def researcher) must match the YAML key (researcher:). Mismatch causes KeyError.

4. Agent.kickoff() — Direct Agent Execution

Use Agent.kickoff() when you need one agent with tools and reasoning, without crew overhead. This is the most common pattern in Flows.

Basic Usage

from crewai import Agent
from crewai_tools import SerperDevTool

researcher = Agent(
    role="Senior Research Analyst",
    goal="Find comprehensive, factual information with source citations",
    backstory="Expert researcher known for thorough, evidence-based analysis.",
    tools=[SerperDevTool()],
    llm="openai/gpt-4o",
)

# Pass a string prompt — the agent reasons, uses tools, and returns a result
result = researcher.kickoff("What are the latest developments in quantum computing?")
print(result.raw)             # str — the agent's full response
print(result.usage_metrics)   # token usage stats

With Structured Output

from pydantic import BaseModel

class ResearchFindings(BaseModel):
    key_trends: list[str]
    sources: list[str]
    confidence: float

result = researcher.kickoff(
    "Research the latest AI agent frameworks",
    response_format=ResearchFindings,
)

# Access via .pydantic (NOT directly — Agent.kickoff wraps the result)
print(result.pydantic.key_trends)    # list[str]
print(result.pydantic.confidence)    # float
print(result.raw)                    # raw string version

Note: Agent.kickoff() returns LiteAgentOutput — access structured output via result.pydantic. This differs from LLM.call() which returns the Pydantic object directly.

With File Inputs

result = researcher.kickoff(
    "Analyze this document and summarize the key findings",
    input_files={"document": FileInput(path="report.pdf")},
)

Async Variant

result = await researcher.kickoff_async(
    "Research quantum computing breakthroughs",
    response_format=ResearchFindings,
)

Agent.kickoff() in Flows (Recommended Pattern)

The most powerful pattern is orchestrating multiple Agent.kickoff() calls inside a Flow. The Flow handles state and sequencing; each agent handles its specific step:

from crewai import Agent
from crewai.flow.flow import Flow, listen, start
from crewai_tools import SerperDevTool, ScrapeWebsiteTool
from pydantic import BaseModel

class ResearchState(BaseModel):
    topic: str = ""
    research: str = ""
    analysis: str = ""
    report: str = ""

class ResearchFlow(Flow[ResearchState]):

    @start()
    def gather_data(self):
        researcher = Agent(
            role="Senior Researcher",
            goal="Find comprehensive data with sources",
            backstory="Expert at finding and validating information.",
            tools=[SerperDevTool(), ScrapeWebsiteTool()],
        )
        result = researcher.kickoff(f"Research: {self.state.topic}")
        self.state.research = result.raw

    @listen(gather_data)
    def analyze(self):
        analyst = Agent(
            role="Data Analyst",
            goal="Extract actionable insights from raw research",
            backstory="Skilled at pattern recognition and synthesis.",
        )
        result = analyst.kickoff(
            f"Analyze this research and extract key insights:\n\n{self.state.research}"
        )
        self.state.analysis = result.raw

    @listen(analyze)
    def write_report(self):
        writer = Agent(
            role="Report Writer",
            goal="Create clear, well-structured reports",
            backstory="Technical writer who makes complex topics accessible.",
        )
        result = writer.kickoff(
            f"Write a comprehensive report from this analysis:\n\n{self.state.analysis}"
        )
        self.state.report = result.raw

flow = ResearchFlow()
flow.kickoff(inputs={"topic": "AI agents"})
print(flow.state.report)

When to use Agent.kickoff() vs Crew.kickoff():

• Use Agent.kickoff() when each step is a distinct agent and the Flow controls sequencing
• Use Crew.kickoff() when multiple agents need to collaborate on related tasks within a single step

5. Specialist vs Generalist Agents

Always prefer specialists. An agent that does one thing well outperforms one that does many things acceptably.

When to Use a Specialist

• Task requires deep domain knowledge
• Output quality matters more than speed
• The task is complex enough to benefit from focused expertise

When a Generalist Is Acceptable

• Simple tasks with clear instructions
• Prototyping where you'll specialize later
• Tasks that truly span multiple domains equally

Specialist Design Pattern

Instead of one "Content Writer" agent, create:

• technical_writer — deep technical accuracy, code examples
• copywriter — persuasive, audience-focused marketing copy
• editor — grammar, consistency, style guide enforcement

Each specialist has a narrow role, specific goal, and backstory that reinforces their expertise.

6. Agent Interaction Patterns

Sequential (Default)

Agents work one after another. Each agent receives prior agents' outputs as context.

Researcher → Writer → Editor

Best for: linear pipelines where each step builds on the last.

Hierarchical

A manager agent delegates and validates. Task assignment is dynamic.

Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    process=Process.hierarchical,
    manager_llm="openai/gpt-4o",
)

Best for: complex workflows where task assignment depends on intermediate results.

Agent-to-Agent Delegation

When allow_delegation=True, an agent can ask another crew agent for help:

lead_researcher = Agent(
    role="Lead Researcher",
    goal="Coordinate research efforts",
    backstory="...",
    allow_delegation=True,  # Can delegate to other agents in the crew
)

The agent will automatically discover other crew members and delegate subtasks as needed.

7. Common Agent Design Mistakes

8. Agent Design Checklist

Before deploying an agent, verify:

• Role is specific and domain-focused (not "Assistant" or "Helper")
• Goal includes desired outcome AND quality standards
• Backstory establishes expertise and working style
• Tools are assigned for any task requiring external data
• No excess tools — 3-5 per agent maximum
• max_iter is tuned for expected task complexity (10-15 for simple, 20-25 for complex)
• max_execution_time is set for production agents to prevent hangs
• Guardrails are configured for critical outputs
• LLM is appropriate for task complexity (don't use GPT-4 for classification)
• Delegation is disabled unless genuinely needed

References

For deeper dives into specific topics, see:

• Custom Tools — building your own tools with @tool decorator and BaseTool subclass
• Memory & Knowledge — memory configuration, knowledge sources, embedder setup, scoping

For related skills:

• getting-started — project scaffolding, choosing the right abstraction, Flow architecture
• design-task — task description/expected_output best practices, guardrails, structured output, dependencies
• ask-docs — query the live CrewAI documentation MCP server for questions not covered by these skills