Getting Started

CrewAI Getting Started & Architecture

How to choose the right abstraction, scaffold a project, and wire everything together.

MANDATORY WORKFLOW — Read This First

NEVER manually create crewAI project files. Always scaffold with the CLI:

crewai create flow <project_name>

This is not optional. Even if you only need one crew, even if you know the file structure by heart — run the CLI first, then modify the generated files. Do NOT write main.py, crew.py, agents.yaml, tasks.yaml, or pyproject.toml by hand from scratch.

Why: The CLI sets up correct imports, directory structure, pyproject.toml config, and boilerplate that is easy to get subtly wrong when done manually. The reference material below teaches you how the pieces work so you can modify scaffolded code, not so you can replace the scaffolding step.

Workflow:

1. Run crewai create flow <name> (use underscores, not hyphens)
2. Edit the generated YAML and Python files to match your use case
3. Run crewai install then crewai run

1. Choosing the Right Abstraction

crewAI has five common abstraction choices. Pick the simplest one that fits your need:

Decision Flowchart

Do you need tools or multi-step reasoning?
├── No  → LLM.call()
└── Yes
    └── Do you need multiple agents collaborating?
        ├── No  → Agent.kickoff()
        └── Yes
            └── Do you need state management, routing, or multiple crews?
                ├── No  → Crew (but still scaffold as a Flow for future-proofing)
                └── Yes → Flow + Crew(s)

Do users send multiple chat messages in one session?
└── Yes → Conversational Flow with handle_turn(message, session_id=...)

Rule of thumb: For any production application, always start with a Flow. You can embed LLM.call(), Agent.kickoff(), or Crew.kickoff() inside Flow steps. This gives you state management, error handling, and room to grow.

For chat applications, start with a conversational Flow rather than trying to make Crew.kickoff() or Flow.kickoff() act like a chat loop. The conversational surface is experimental, but it is the intended API for multi-turn sessions: call flow.handle_turn(message, session_id=...) for every user line, or flow.chat() for a local terminal REPL. Official guide: https://docs.crewai.com/en/guides/flows/conversational-flows.

2. LLM.call() — Direct LLM Invocation

Use for simple, single-turn tasks where you don't need tools or agent reasoning.

from crewai import LLM
from pydantic import BaseModel

class EmailFields(BaseModel):
    sender: str
    subject: str
    urgency: str

llm = LLM(model="openai/gpt-4o")

# Without response_format — returns a string
raw = llm.call(messages=[{"role": "user", "content": "Summarize this text..."}])
print(raw)  # str

# With response_format — returns the Pydantic object directly
result = llm.call(
    messages=[{"role": "user", "content": f"Extract fields from this email: {email_text}"}],
    response_format=EmailFields
)
print(result.sender)   # str — access Pydantic fields directly
print(result.urgency)  # str

When NOT to use: If you need tools, multi-step reasoning, or retries — use an Agent instead.

3. Agent.kickoff() — Single Agent Execution

Use when you need one agent with tools and reasoning, but don't need multi-agent coordination.

from crewai import Agent
from crewai_tools import SerperDevTool
from pydantic import BaseModel

class ResearchFindings(BaseModel):
    main_points: list[str]
    key_technologies: list[str]

researcher = Agent(
    role="AI Researcher",
    goal="Research the latest AI developments",
    backstory="Expert AI researcher with deep technical knowledge.",
    llm="openai/gpt-4o",       # Optional: defaults to OPENAI_MODEL_NAME env var or "gpt-4"
    tools=[SerperDevTool()],
)

# Unstructured output
result = researcher.kickoff("What are the latest LLM developments?")
print(result.raw)            # str
print(result.usage_metrics)  # token usage

# Structured output with response_format
result = researcher.kickoff(
    "Summarize latest AI developments",
    response_format=ResearchFindings,
)
print(result.pydantic.main_points)

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

When NOT to use: If you need multiple agents passing context to each other — use a Crew.

4. CLI Scaffold Reference

As stated above: NEVER skip crewai create flow. This section documents what the CLI generates so you know what to modify — not so you can recreate it by hand.

crewai create flow my_project

Warning: Always use underscores in project names, not hyphens. crewai create flow my-project creates a directory that is not a valid Python identifier, causing ModuleNotFoundError on import. Use my_project instead.

This generates:

my_project/
├── src/my_project/
│   ├── crews/
│   │   └── my_crew/
│   │       ├── config/
│   │       │   ├── agents.yaml    # Agent definitions (role, goal, backstory)
│   │       │   └── tasks.yaml     # Task definitions (description, expected_output)
│   │       └── my_crew.py         # Crew class with @CrewBase
│   ├── tools/
│   │   └── custom_tool.py
│   ├── main.py                    # Flow class with @start/@listen
│   └── ...
├── .env                           # API keys (OPENAI_API_KEY, etc.)
└── pyproject.toml

Do not use crewai create crew unless you are certain you will never need routing, state, or multiple crews. Prefer crewai create flow as the default.

5. YAML Configuration (agents.yaml & tasks.yaml)

The scaffold uses YAML files for agent and task definitions. This separates configuration from code and supports {variable} interpolation.

agents.yaml

researcher:
  role: >
    {topic} Senior Data Researcher
  goal: >
    Uncover cutting-edge developments in {topic}
  backstory: >
    You're a seasoned researcher with a knack for uncovering
    the latest developments in {topic}.
  # Optional overrides:
  # llm: openai/gpt-4o
  # max_iter: 20
  # max_rpm: 10

reporting_analyst:
  role: >
    {topic} Reporting Analyst
  goal: >
    Create detailed reports based on {topic} research findings
  backstory: >
    You're a meticulous analyst known for turning complex data
    into clear, actionable reports.

tasks.yaml

research_task:
  description: >
    Conduct thorough research about {topic}.
    Identify key trends, breakthrough technologies,
    and potential industry impacts.
  expected_output: >
    A detailed report with analysis of the top 5
    developments in {topic}, with sources and implications.
  agent: researcher

reporting_task:
  description: >
    Review the research and create a comprehensive report about {topic}.
  expected_output: >
    A polished report formatted in markdown with sections
    for each key finding.
  agent: reporting_analyst
  output_file: output/report.md

Key rules:

• {variable} placeholders are replaced at runtime via crew.kickoff(inputs={...})
• expected_output is always a string (never a Pydantic class name)
• agent value must match an agent key in agents.yaml
• In Process.sequential, each task auto-receives all prior task outputs as context
• For non-sequential deps, use context=[other_task] to explicitly pass output

6. Wiring It Together — crew.py

The @CrewBase decorator auto-loads YAML config files and collects @agent and @task methods.

from crewai import Agent, Crew, Process, Task
from crewai.project import CrewBase, agent, crew, task
from crewai_tools import SerperDevTool

@CrewBase
class ResearchCrew:
    """Research and reporting crew."""

    agents_config = "config/agents.yaml"
    tasks_config = "config/tasks.yaml"

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

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

    @task
    def research_task(self) -> Task:
        return Task(config=self.tasks_config["research_task"])

    @task
    def reporting_task(self) -> Task:
        return Task(
            config=self.tasks_config["reporting_task"],
            context=[self.research_task()],  # Explicit dependency (optional in sequential)
            output_file="output/report.md",
        )

    @crew
    def crew(self) -> Crew:
        return Crew(
            agents=self.agents,  # auto-collected by @agent
            tasks=self.tasks,    # auto-collected by @task
            process=Process.sequential,
            verbose=True,
        )

Important: Method names must match YAML keys. def researcher(self) maps to the researcher: key in agents.yaml.

7. Flows — The Production Foundation

Flows are the recommended way to build production crewAI applications. They provide state management, conditional routing, human-in-the-loop, and persistence — wrapping crews, agents, and LLM calls into a coherent workflow.

Basic Flow — main.py

from crewai.flow.flow import Flow, listen, start
from pydantic import BaseModel
from .crews.research_crew.research_crew import ResearchCrew

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

class ResearchFlow(Flow[ResearchState]):

    @start()
    def begin(self):
        print(f"Starting research on: {self.state.topic}")

    @listen(begin)
    def run_research(self):
        result = ResearchCrew().crew().kickoff(
            inputs={"topic": self.state.topic}
        )
        self.state.report = result.raw

def kickoff():
    flow = ResearchFlow()
    flow.kickoff(inputs={"topic": "AI Agents"})

if __name__ == "__main__":
    kickoff()

Key points:

• flow.kickoff(inputs={"topic": "AI Agents"}) populates self.state.topic (keys must match Pydantic field names). The YAML {variable} substitution happens later, when you call crew.kickoff(inputs={"topic": self.state.topic}) inside a Flow step. The chain is: flow inputs → state → crew inputs → YAML substitution.
• Each @listen method runs after its dependency completes
• State persists across all Flow steps — use it to pass data between crews

State Management — Structured vs Unstructured

Structured (recommended for production):

from pydantic import BaseModel

class MyState(BaseModel):
    topic: str = ""
    research: str = ""
    draft: str = ""
    approved: bool = False

class MyFlow(Flow[MyState]):
    ...

Unstructured (quick prototyping):

class MyFlow(Flow):  # No type parameter — state is a dict
    @start()
    def begin(self):
        self.state["topic"] = "AI"  # dict-style access

Use structured state for type safety, IDE autocompletion, and validation. Use unstructured only for throwaway prototypes.

Using Agent.kickoff() Inside Flows (Common Pattern)

Many production Flows skip Crews entirely and orchestrate individual agents via Agent.kickoff(). This gives you fine-grained control — each Flow step calls a specific agent, passes state, and stores the result. The Flow handles orchestration; agents handle reasoning.

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

class ResearchState(BaseModel):
    query: str = ""
    raw_research: str = ""
    analysis: str = ""
    report: str = ""

class DeepResearchFlow(Flow[ResearchState]):

    @start()
    def gather_research(self):
        """Agent with tools does the actual searching."""
        researcher = Agent(
            role="Senior Research Analyst",
            goal="Find comprehensive, factual information about the given topic",
            backstory="You're an expert researcher who always cites sources and flags uncertainty.",
            tools=[SerperDevTool(), ScrapeWebsiteTool()],
            llm="openai/gpt-4o",
        )
        result = researcher.kickoff(
            f"Research this topic thoroughly: {self.state.query}"
        )
        self.state.raw_research = result.raw

    @listen(gather_research)
    def analyze_findings(self):
        """A different agent analyzes the raw research — no tools needed."""
        analyst = Agent(
            role="Data Analyst",
            goal="Extract key insights, patterns, and actionable recommendations",
            backstory="You turn raw data into clear, structured analysis.",
            llm="openai/gpt-4o",
        )
        result = analyst.kickoff(
            f"Analyze these research findings and extract key insights:\n\n{self.state.raw_research}"
        )
        self.state.analysis = result.raw

    @listen(analyze_findings)
    def write_report(self):
        """A writer agent produces the final deliverable."""
        writer = Agent(
            role="Technical Writer",
            goal="Produce clear, actionable reports for non-technical readers",
            backstory="You specialize in making complex information accessible.",
            llm="openai/gpt-4o",
        )
        result = writer.kickoff(
            f"Write a comprehensive report based on this analysis:\n\n{self.state.analysis}"
        )
        self.state.report = result.raw

Why this pattern works well:

• Each agent is purpose-built for its step — narrow role, specific tools
• The Flow manages state and sequencing — no crew overhead
• Easy to add routing, human review, or retry logic between steps
• You can mix Agent.kickoff(), LLM.call(), and Crew.kickoff() freely

When to use Agent.kickoff() vs Crew.kickoff() in a Flow:

Agent.kickoff() with Structured Output in Flows

Combine response_format with state for typed data flow between agents:

class Insights(BaseModel):
    key_points: list[str]
    recommendations: list[str]
    confidence: float

class AnalysisFlow(Flow[AnalysisState]):

    @start()
    def research(self):
        researcher = Agent(role="Researcher", goal="...", backstory="...", tools=[SerperDevTool()])
        result = researcher.kickoff(
            f"Research {self.state.topic}",
            response_format=Insights,
        )
        # result.pydantic gives you the typed Insights object
        self.state.key_points = result.pydantic.key_points
        self.state.recommendations = result.pydantic.recommendations

Mixing Abstractions in a Flow

A Flow can combine all crewAI abstractions in a single workflow:

class ProductFlow(Flow[ProductState]):

    @start()
    def classify_request(self):
        # LLM.call() for simple classification
        llm = LLM(model="openai/gpt-4o")
        self.state.category = llm.call(
            messages=[{"role": "user", "content": f"Classify: {self.state.request}"}],
            response_format=Category
        ).category

    @router(classify_request)
    def route_by_category(self):
        if self.state.category == "simple":
            return "quick_answer"
        return "deep_research"

    @listen("quick_answer")
    def handle_simple(self):
        # Agent.kickoff() for single-agent work
        agent = Agent(role="Helper", goal="Answer quickly", backstory="...")
        result = agent.kickoff(self.state.request)
        self.state.answer = result.raw

    @listen("deep_research")
    def handle_complex(self):
        # Crew.kickoff() for multi-agent collaboration
        result = ResearchCrew().crew().kickoff(
            inputs={"topic": self.state.request}
        )
        self.state.answer = result.raw

Flow Routing with @router

Use @router for conditional branching — return a string label, and @listen("label") binds to branches:

from crewai.flow.flow import Flow, listen, router, start, or_

class QualityFlow(Flow[QAState]):

    @start()
    def generate_content(self):
        result = WriterCrew().crew().kickoff(inputs={"topic": self.state.topic})
        self.state.draft = result.raw

    @router(generate_content)
    def check_quality(self):
        llm = LLM(model="openai/gpt-4o")
        score = llm.call(
            messages=[{"role": "user", "content": f"Rate 1-10: {self.state.draft}"}],
            response_format=QualityScore
        )
        if score.rating >= 7:
            return "approved"
        return "needs_revision"

    @listen("approved")
    def publish(self):
        self.state.published = True

    @listen("needs_revision")
    def revise(self):
        result = EditorCrew().crew().kickoff(
            inputs={"draft": self.state.draft}
        )
        self.state.draft = result.raw

Converging Branches with or_() and and_()

from crewai.flow.flow import Flow, listen, start, or_, and_

class ParallelFlow(Flow[MyState]):

    @start()
    def fetch_data_a(self):
        ...

    @start()
    def fetch_data_b(self):
        ...

    # Runs when BOTH fetches complete
    @listen(and_(fetch_data_a, fetch_data_b))
    def merge_results(self):
        ...

    # Runs when EITHER source provides data
    @listen(or_(fetch_data_a, fetch_data_b))
    def process_first_available(self):
        ...

Flow Persistence with @persist

For long-running workflows that need to survive restarts:

from crewai.flow.flow import Flow, start, listen, persist
from crewai.flow.persistence import SQLiteFlowPersistence

@persist(SQLiteFlowPersistence())  # Class-level: persists all methods
class LongRunningFlow(Flow[MyState]):

    @start()
    def step_one(self):
        self.state.data = "processed"

    @listen(step_one)
    def step_two(self):
        # If the process crashes here, restarting with the same
        # state ID will resume from after step_one
        ...

Conversational Flows with handle_turn() (Experimental)

Use a conversational Flow when the product is a chat session: support assistants, routed research helpers, onboarding wizards, or any UI where the same user sends multiple turns.

Core model:

• Each user message is a new Flow run with the same session id
• handle_turn(message, session_id=...) appends the user line to state.messages, resets per-turn execution tracking, and calls kickoff(inputs={"id": session_id}) internally
• Flow.kickoff() does not accept user_message= or session_id= keyword args
• Route chat turns with route_turn() plus @listen("ROUTE") handlers
• Call append_assistant_message(reply) in handlers so the next turn sees assistant history
• Wrap owned loops in try/finally and call finalize_session_traces(); flow.chat() does this for local REPLs

from uuid import uuid4

from crewai import Agent, Flow
from crewai.flow import listen
from crewai.experimental.conversational import (
    ConversationConfig,
    ConversationState,
)


@ConversationConfig(defer_trace_finalization=True)
class SupportFlow(Flow[ConversationState]):
    conversational = True

    def research_agent(self) -> Agent:
        return Agent(
            role="Support Research Specialist",
            goal="Answer the user's current research question with accurate sources.",
            backstory="You are precise, evidence-driven, and explicit about uncertainty.",
            tools=[...],
        )

    def route_turn(self, context):
        message = (self.state.current_user_message or "").lower()
        if "docs" in message or "crewai" in message:
            return "CREWAI_DOCS"
        if "research" in message or "search" in message:
            return "RESEARCH"
        return "converse"

    @listen("CREWAI_DOCS")
    def handle_docs(self):
        """Look up CrewAI documentation for framework/API questions."""
        reply = "I would query the CrewAI docs here."
        self.append_assistant_message(reply)
        return reply

    @listen("RESEARCH")
    def handle_research(self):
        """Fresh research, current lookups, and tool-backed investigation."""
        result = self.research_agent().kickoff(self.state.current_user_message)
        reply = result.raw
        self.append_assistant_message(reply)
        return reply


flow = SupportFlow()
session_id = str(uuid4())

try:
    flow.handle_turn("What can you do?", session_id=session_id)
    flow.handle_turn("Check the CrewAI docs for flows.", session_id=session_id)
finally:
    flow.finalize_session_traces()

Use RouterConfig when you want LLM-driven routing. The router catalog is auto-built from @listen("ROUTE") handlers and their docstrings, so do not duplicate the route list in the router prompt.

See Conversational Flows for the full lifecycle, routing, persistence, and trace guidance.

Human-in-the-Loop with @human_feedback

from crewai.flow.flow import Flow, start, listen, router
from crewai.flow.human_feedback import human_feedback

class ApprovalFlow(Flow[ReviewState]):

    @start()
    def generate_draft(self):
        result = WriterCrew().crew().kickoff(inputs={"topic": self.state.topic})
        self.state.draft = result.raw

    @human_feedback(
        message="Review the draft and provide feedback",
        emit=["approved", "needs_revision"],
        llm="openai/gpt-4o",
        default_outcome="approved"
    )
    @listen(generate_draft)
    def review_step(self):
        return self.state.draft

    @listen("approved")
    def publish(self):
        ...

    @listen("needs_revision")
    def revise(self):
        feedback = self.last_human_feedback
        # Use feedback.feedback_text for revision
        ...

Flow Visualization

flow = MyFlow()
flow.plot()             # Display in notebook
flow.plot("my_flow")    # Save as my_flow.png

8. Variable Interpolation with inputs

The {variable} pattern is how you make crews reusable.

# Variables flow through: kickoff → YAML templates → agent/task prompts
crew.kickoff(inputs={
    "topic": "AI Agents",
    "current_year": "2025",
    "target_audience": "developers",
})

In YAML, {topic} and {current_year} get replaced:

research_task:
  description: >
    Research {topic} trends for {current_year},
    targeting {target_audience}.

Common mistakes:

• Forgetting to pass a variable that's referenced in YAML → results in literal {variable} in the prompt
• Using Jinja2 syntax {{ }} instead of single-brace { } → crewAI uses single braces
• Passing variables that don't match any YAML placeholder → silently ignored

9. Running Your Project

# Install dependencies
crewai install

# Run the flow
crewai run

Or run directly:

cd my_project
uv run src/my_project/main.py

10. Quick Diagnostic Checklist

References

For deeper dives into specific topics, see:

• Flow Routing, Persistence, Streaming & Human Feedback — complete @router, or_(), and_(), @persist, streaming, and @human_feedback patterns
• Conversational Flows — experimental multi-turn Flow API with handle_turn(), chat(), ConversationConfig, router behavior, persistence, and tracing
• MCP Servers — prefer official MCP servers over native tools; setup, DSL integration, and known official servers
• Tools Catalog — all 80+ built-in tools with imports, env vars, and common combos (use as fallback when no MCP server exists)

For related skills:

• design-agent — agent Role-Goal-Backstory framework, parameter tuning, tool assignment, memory & knowledge configuration
• 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