Flows - CrewAI

Overview

CrewAI Flows is a powerful feature designed to streamline the creation and management of AI workflows. Flows allow developers to combine and coordinate coding tasks and Crews efficiently, providing a robust framework for building sophisticated AI automations. Flows allow you to create structured, event-driven workflows. They provide a seamless way to connect multiple tasks, manage state, and control the flow of execution in your AI applications. With Flows, you can easily design and implement multi-step processes that leverage the full potential of CrewAI’s capabilities.

Simplified Workflow Creation: Easily chain together multiple Crews and tasks to create complex AI workflows.
State Management: Flows make it super easy to manage and share state between different tasks in your workflow.
Event-Driven Architecture: Built on an event-driven model, allowing for dynamic and responsive workflows.
Flexible Control Flow: Implement conditional logic, loops, and branching within your workflows.

Getting Started

Let’s create a simple Flow where you will use OpenAI to generate a random city in one task and then use that city to generate a fun fact in another task.

Code

from crewai.flow.flow import Flow, listen, start
from dotenv import load_dotenv
from litellm import completion

load_dotenv()

class ExampleFlow(Flow):
    model = "gpt-4o-mini"

    @start()
    def generate_city(self):
        print("Starting flow")
        # Each flow state automatically gets a unique ID
        print(f"Flow State ID: {self.state['id']}")

        response = completion(
            model=self.model,
            messages=[
                {
                    "role": "user",
                    "content": "Return the name of a random city in the world.",
                },
            ],
        )

        random_city = response["choices"][0]["message"]["content"]
        # Store the city in our state
        self.state["city"] = random_city
        print(f"Random City: {random_city}")

        return random_city

    @listen(generate_city)
    def generate_fun_fact(self, random_city):
        response = completion(
            model=self.model,
            messages=[
                {
                    "role": "user",
                    "content": f"Tell me a fun fact about {random_city}",
                },
            ],
        )

        fun_fact = response["choices"][0]["message"]["content"]
        # Store the fun fact in our state
        self.state["fun_fact"] = fun_fact
        return fun_fact



flow = ExampleFlow()
flow.plot()
result = flow.kickoff()

print(f"Generated fun fact: {result}")

In the above example, we have created a simple Flow that generates a random city using OpenAI and then generates a fun fact about that city. The Flow consists of two tasks: generate_city and generate_fun_fact. The generate_city task is the starting point of the Flow, and the generate_fun_fact task listens for the output of the generate_city task. Each Flow instance automatically receives a unique identifier (UUID) in its state, which helps track and manage flow executions. The state can also store additional data (like the generated city and fun fact) that persists throughout the flow’s execution. When you run the Flow, it will:

Generate a unique ID for the flow state
Generate a random city and store it in the state
Generate a fun fact about that city and store it in the state
Print the results to the console

The state’s unique ID and stored data can be useful for tracking flow executions and maintaining context between tasks. Note: Ensure you have set up your .env file to store your OPENAI_API_KEY. This key is necessary for authenticating requests to the OpenAI API.

@start()

The @start() decorator marks entry points for a Flow. You can:

Declare multiple unconditional starts: @start()
Gate a start on a prior method or router label: @start("method_or_label")
Provide a callable condition to control when a start should fire

All satisfied @start() methods will execute (often in parallel) when the Flow begins or resumes.

@listen()

The @listen() decorator is used to mark a method as a listener for the output of another task in the Flow. The method decorated with @listen() will be executed when the specified task emits an output. The method can access the output of the task it is listening to as an argument.

Usage

The @listen() decorator can be used in several ways:

Listening to a Method by Name: You can pass the name of the method you want to listen to as a string. When that method completes, the listener method will be triggered.
Code
@listen("generate_city") def generate_fun_fact(self, random_city): # Implementation
Listening to a Method Directly: You can pass the method itself. When that method completes, the listener method will be triggered.
Code
@listen(generate_city) def generate_fun_fact(self, random_city): # Implementation

Flow Output

Accessing and handling the output of a Flow is essential for integrating your AI workflows into larger applications or systems. CrewAI Flows provide straightforward mechanisms to retrieve the final output, access intermediate results, and manage the overall state of your Flow.

Retrieving the Final Output

When you run a Flow, the final output is determined by the last method that completes. The kickoff() method returns the output of this final method. Here’s how you can access the final output:

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

class OutputExampleFlow(Flow):
    @start()
    def first_method(self):
        return "Output from first_method"

    @listen(first_method)
    def second_method(self, first_output):
        return f"Second method received: {first_output}"


flow = OutputExampleFlow()
flow.plot("my_flow_plot")
final_output = flow.kickoff()

print("---- Final Output ----")
print(final_output)

---- Final Output ----
Second method received: Output from first_method

In this example, the second_method is the last method to complete, so its output will be the final output of the Flow. The kickoff() method will return the final output, which is then printed to the console. The plot() method will generate the HTML file, which will help you understand the flow.

Accessing and Updating State

In addition to retrieving the final output, you can also access and update the state within your Flow. The state can be used to store and share data between different methods in the Flow. After the Flow has run, you can access the state to retrieve any information that was added or updated during the execution. Here’s an example of how to update and access the state:

from crewai.flow.flow import Flow, listen, start
from pydantic import BaseModel

class ExampleState(BaseModel):
    counter: int = 0
    message: str = ""

class StateExampleFlow(Flow[ExampleState]):

    @start()
    def first_method(self):
        self.state.message = "Hello from first_method"
        self.state.counter += 1

    @listen(first_method)
    def second_method(self):
        self.state.message += " - updated by second_method"
        self.state.counter += 1
        return self.state.message

flow = StateExampleFlow()
flow.plot("my_flow_plot")
final_output = flow.kickoff()
print(f"Final Output: {final_output}")
print("Final State:")
print(flow.state)

Final Output: Hello from first_method - updated by second_method
Final State:
counter=2 message='Hello from first_method - updated by second_method'

In this example, the state is updated by both first_method and second_method. After the Flow has run, you can access the final state to see the updates made by these methods. By ensuring that the final method’s output is returned and providing access to the state, CrewAI Flows make it easy to integrate the results of your AI workflows into larger applications or systems, while also maintaining and accessing the state throughout the Flow’s execution.

Flow Usage Metrics

After a Flow execution completes, you can access the usage_metrics property to view aggregated token usage across every LLM call made during the run — including calls from every Crew the Flow orchestrated, calls inside Agent tools, and bare LLM.call(...) invocations from Flow methods. This is the SDK-side equivalent of the totals shown in the CrewAI Enterprise UI.

Code

from crewai import LLM
from crewai.flow.flow import Flow, listen, start

class UsageMetricsFlow(Flow):
    @start()
    def run_first_crew(self):
        self.state.first_result = FirstCrew().crew().kickoff()

    @listen(run_first_crew)
    def call_llm_directly(self):
        # Bare LLM call — still counted by flow.usage_metrics
        llm = LLM(model="openai/gpt-4o-mini")
        self.state.summary = llm.call("Summarize the key takeaways.")

    @listen(call_llm_directly)
    def run_second_crew(self):
        self.state.second_result = SecondCrew().crew().kickoff()

flow = UsageMetricsFlow()
flow.kickoff()

print(flow.usage_metrics)
# UsageMetrics(total_tokens=8579, prompt_tokens=6210, completion_tokens=2369,
#              cached_prompt_tokens=0, reasoning_tokens=0,
#              cache_creation_tokens=0, successful_requests=5)

flow.usage_metrics is not the same as flow.kickoff().token_usage. The latter returns the CrewOutput.token_usage of the last @listen method that returned a CrewOutput, which means it only reflects the final Crew and ignores prior Crews and bare LLM.call(...) invocations entirely. Use flow.usage_metrics whenever you need the full token rollup for the Flow execution.

Each entry in the returned UsageMetrics is the sum across all LLM calls made within a single flow.kickoff() invocation. Counters reset on the next kickoff() call (or on each iteration of kickoff_for_each), so successive runs don’t double-count. The property is safe to read at any point after kickoff() completes; reading it during execution returns the partial total accumulated so far.

Flow State Management

Managing state effectively is crucial for building reliable and maintainable AI workflows. CrewAI Flows provides robust mechanisms for both unstructured and structured state management, allowing developers to choose the approach that best fits their application’s needs.

Unstructured State Management

In unstructured state management, all state is stored in the state attribute of the Flow class. This approach offers flexibility, enabling developers to add or modify state attributes on the fly without defining a strict schema. Even with unstructured states, CrewAI Flows automatically generates and maintains a unique identifier (UUID) for each state instance.

Code

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

class UnstructuredExampleFlow(Flow):

    @start()
    def first_method(self):
        # The state automatically includes an 'id' field
        print(f"State ID: {self.state['id']}")
        self.state['counter'] = 0
        self.state['message'] = "Hello from structured flow"

    @listen(first_method)
    def second_method(self):
        self.state['counter'] += 1
        self.state['message'] += " - updated"

    @listen(second_method)
    def third_method(self):
        self.state['counter'] += 1
        self.state['message'] += " - updated again"

        print(f"State after third_method: {self.state}")


flow = UnstructuredExampleFlow()
flow.plot("my_flow_plot")
flow.kickoff()

Note: The id field is automatically generated and preserved throughout the flow’s execution. You don’t need to manage or set it manually, and it will be maintained even when updating the state with new data. Key Points:

Flexibility: You can dynamically add attributes to self.state without predefined constraints.
Simplicity: Ideal for straightforward workflows where state structure is minimal or varies significantly.

Structured State Management

Structured state management leverages predefined schemas to ensure consistency and type safety across the workflow. By using models like Pydantic’s BaseModel, developers can define the exact shape of the state, enabling better validation and auto-completion in development environments. Each state in CrewAI Flows automatically receives a unique identifier (UUID) to help track and manage state instances. This ID is automatically generated and managed by the Flow system.

Code

from crewai.flow.flow import Flow, listen, start
from pydantic import BaseModel


class ExampleState(BaseModel):
    # Note: 'id' field is automatically added to all states
    counter: int = 0
    message: str = ""


class StructuredExampleFlow(Flow[ExampleState]):

    @start()
    def first_method(self):
        # Access the auto-generated ID if needed
        print(f"State ID: {self.state.id}")
        self.state.message = "Hello from structured flow"

    @listen(first_method)
    def second_method(self):
        self.state.counter += 1
        self.state.message += " - updated"

    @listen(second_method)
    def third_method(self):
        self.state.counter += 1
        self.state.message += " - updated again"

        print(f"State after third_method: {self.state}")


flow = StructuredExampleFlow()
flow.kickoff()

Key Points:

Defined Schema: ExampleState clearly outlines the state structure, enhancing code readability and maintainability.
Type Safety: Leveraging Pydantic ensures that state attributes adhere to the specified types, reducing runtime errors.
Auto-Completion: IDEs can provide better auto-completion and error checking based on the defined state model.

Choosing Between Unstructured and Structured State Management

Use Unstructured State Management when:
- The workflow’s state is simple or highly dynamic.
- Flexibility is prioritized over strict state definitions.
- Rapid prototyping is required without the overhead of defining schemas.
Use Structured State Management when:
- The workflow requires a well-defined and consistent state structure.
- Type safety and validation are important for your application’s reliability.
- You want to leverage IDE features like auto-completion and type checking for better developer experience.

By providing both unstructured and structured state management options, CrewAI Flows empowers developers to build AI workflows that are both flexible and robust, catering to a wide range of application requirements.

Flow Persistence

The @persist decorator enables automatic state persistence in CrewAI Flows, allowing you to maintain flow state across restarts or different workflow executions. This decorator can be applied at either the class level or method level, providing flexibility in how you manage state persistence.

Class-Level Persistence

When applied at the class level, the @persist decorator automatically persists all flow method states:

@persist  # Using SQLiteFlowPersistence by default
class MyFlow(Flow[MyState]):
    @start()
    def initialize_flow(self):
        # This method will automatically have its state persisted
        self.state.counter = 1
        print("Initialized flow. State ID:", self.state.id)

    @listen(initialize_flow)
    def next_step(self):
        # The state (including self.state.id) is automatically reloaded
        self.state.counter += 1
        print("Flow state is persisted. Counter:", self.state.counter)

Method-Level Persistence

For more granular control, you can apply @persist to specific methods:

class AnotherFlow(Flow[dict]):
    @persist  # Persists only this method's state
    @start()
    def begin(self):
        if "runs" not in self.state:
            self.state["runs"] = 0
        self.state["runs"] += 1
        print("Method-level persisted runs:", self.state["runs"])

Forking Persisted State

@persist supports two distinct hydration modes on kickoff / kickoff_async:

kickoff(inputs={"id": <uuid>}) — resume: load the latest snapshot for the supplied UUID and continue writing under the same flow_uuid. The history extends.
kickoff(restore_from_state_id=<uuid>) — fork: load the latest snapshot for the supplied UUID, hydrate the new run’s state from it, and assign a fresh state.id (auto-generated, or inputs["id"] if pinned). The new run’s @persist writes land under the new state.id; the source flow’s history is preserved.

from crewai.flow.flow import Flow, start
from crewai.flow.persistence import persist
from pydantic import BaseModel

class CounterState(BaseModel):
    id: str = ""
    counter: int = 0

@persist
class CounterFlow(Flow[CounterState]):
    @start()
    def step(self):
        self.state.counter += 1
        print(f"[id={self.state.id}] counter={self.state.counter}")

# Run 1: fresh state, counter 0 -> 1, persisted under flow_1.state.id
flow_1 = CounterFlow()
flow_1.kickoff()

# Fork: hydrate from flow_1's latest snapshot, but use a NEW state.id
flow_2 = CounterFlow()
flow_2.kickoff(restore_from_state_id=flow_1.state.id)
# flow_2.state.counter starts at 1 (hydrated), then step() bumps it to 2.
# flow_2.state.id != flow_1.state.id; flow_1's history is unchanged.

If the supplied restore_from_state_id does not match any persisted state, the kickoff falls back silently — same as the existing inputs["id"] resume not-found behavior. Combining restore_from_state_id with from_checkpoint raises a ValueError; pick one hydration source. Pinning inputs["id"] while forking shares a persistence key with another flow — usually you want only restore_from_state_id.

How It Works

Unique State Identification
- Each flow state automatically receives a unique UUID
- The ID is preserved across state updates and method calls
- Supports both structured (Pydantic BaseModel) and unstructured (dictionary) states
Default SQLite Backend
- SQLiteFlowPersistence is the default storage backend
- States are automatically saved to a local SQLite database
- Robust error handling ensures clear messages if database operations fail
Error Handling
- Comprehensive error messages for database operations
- Automatic state validation during save and load
- Clear feedback when persistence operations encounter issues

Important Considerations

State Types: Both structured (Pydantic BaseModel) and unstructured (dictionary) states are supported
Automatic ID: The id field is automatically added if not present
State Recovery: Failed or restarted flows can automatically reload their previous state
Custom Implementation: You can provide your own FlowPersistence implementation for specialized storage needs

Technical Advantages

Precise Control Through Low-Level Access
- Direct access to persistence operations for advanced use cases
- Fine-grained control via method-level persistence decorators
- Built-in state inspection and debugging capabilities
- Full visibility into state changes and persistence operations
Enhanced Reliability
- Automatic state recovery after system failures or restarts
- Transaction-based state updates for data integrity
- Comprehensive error handling with clear error messages
- Robust validation during state save and load operations
Extensible Architecture
- Customizable persistence backend through FlowPersistence interface
- Support for specialized storage solutions beyond SQLite
- Compatible with both structured (Pydantic) and unstructured (dict) states
- Seamless integration with existing CrewAI flow patterns

The persistence system’s architecture emphasizes technical precision and customization options, allowing developers to maintain full control over state management while benefiting from built-in reliability features.

Flow Control

Conditional Logic: `or`

The or_ function in Flows allows you to listen to multiple methods and trigger the listener method when any of the specified methods emit an output.

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

class OrExampleFlow(Flow):

    @start()
    def start_method(self):
        return "Hello from the start method"

    @listen(start_method)
    def second_method(self):
        return "Hello from the second method"

    @listen(or_(start_method, second_method))
    def logger(self, result):
        print(f"Logger: {result}")



flow = OrExampleFlow()
flow.plot("my_flow_plot")
flow.kickoff()

Logger: Hello from the start method
Logger: Hello from the second method

When you run this Flow, the logger method will be triggered by the output of either the start_method or the second_method. The or_ function is used to listen to multiple methods and trigger the listener method when any of the specified methods emit an output.

Conditional Logic: `and`

The and_ function in Flows allows you to listen to multiple methods and trigger the listener method only when all the specified methods emit an output.

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

class AndExampleFlow(Flow):

    @start()
    def start_method(self):
        self.state["greeting"] = "Hello from the start method"

    @listen(start_method)
    def second_method(self):
        self.state["joke"] = "What do computers eat? Microchips."

    @listen(and_(start_method, second_method))
    def logger(self):
        print("---- Logger ----")
        print(self.state)

flow = AndExampleFlow()
flow.plot()
flow.kickoff()

---- Logger ----
{'greeting': 'Hello from the start method', 'joke': 'What do computers eat? Microchips.'}

When you run this Flow, the logger method will be triggered only when both the start_method and the second_method emit an output. The and_ function is used to listen to multiple methods and trigger the listener method only when all the specified methods emit an output.

Router

The @router() decorator in Flows allows you to define conditional routing logic based on the output of a method. You can specify different routes based on the output of the method, allowing you to control the flow of execution dynamically.

import random
from crewai.flow.flow import Flow, listen, router, start
from pydantic import BaseModel

class ExampleState(BaseModel):
    success_flag: bool = False

class RouterFlow(Flow[ExampleState]):

    @start()
    def start_method(self):
        print("Starting the structured flow")
        random_boolean = random.choice([True, False])
        self.state.success_flag = random_boolean

    @router(start_method)
    def second_method(self):
        if self.state.success_flag:
            return "success"
        else:
            return "failed"

    @listen("success")
    def third_method(self):
        print("Third method running")

    @listen("failed")
    def fourth_method(self):
        print("Fourth method running")


flow = RouterFlow()
flow.plot("my_flow_plot")
flow.kickoff()

Starting the structured flow
Third method running
Fourth method running

In the above example, the start_method generates a random boolean value and sets it in the state. The second_method uses the @router() decorator to define conditional routing logic based on the value of the boolean. If the boolean is True, the method returns "success", and if it is False, the method returns "failed". The third_method and fourth_method listen to the output of the second_method and execute based on the returned value. When you run this Flow, the output will change based on the random boolean value generated by the start_method.

Human in the Loop (human feedback)

The @human_feedback decorator requires CrewAI version 1.8.0 or higher.

The @human_feedback decorator enables human-in-the-loop workflows by pausing flow execution to collect feedback from a human. This is useful for approval gates, quality review, and decision points that require human judgment.

Code

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

class ReviewFlow(Flow):
    @start()
    @human_feedback(
        message="Do you approve this content?",
        emit=["approved", "rejected", "needs_revision"],
        llm="gpt-4o-mini",
        default_outcome="needs_revision",
    )
    def generate_content(self):
        return "Content to be reviewed..."

    @listen("approved")
    def on_approval(self, result: HumanFeedbackResult):
        print(f"Approved! Feedback: {result.feedback}")

    @listen("rejected")
    def on_rejection(self, result: HumanFeedbackResult):
        print(f"Rejected. Reason: {result.feedback}")

When emit is specified, the human’s free-form feedback is interpreted by an LLM and collapsed into one of the specified outcomes, which then triggers the corresponding @listen decorator. You can also use @human_feedback without routing to simply collect feedback:

Code

@start()
@human_feedback(message="Any comments on this output?")
def my_method(self):
    return "Output for review"

@listen(my_method)
def next_step(self, result: HumanFeedbackResult):
    # Access feedback via result.feedback
    # Access original output via result.output
    pass

Access all feedback collected during a flow via self.last_human_feedback (most recent) or self.human_feedback_history (all feedback as a list). For a complete guide on human feedback in flows, including async/non-blocking feedback with custom providers (Slack, webhooks, etc.), see Human Feedback in Flows.

Adding Agents to Flows

Agents can be seamlessly integrated into your flows, providing a lightweight alternative to full Crews when you need simpler, focused task execution. Here’s an example of how to use an Agent within a flow to perform market research:

import asyncio
from typing import Any, Dict, List

from crewai_tools import SerperDevTool
from pydantic import BaseModel, Field

from crewai.agent import Agent
from crewai.flow.flow import Flow, listen, start


# Define a structured output format
class MarketAnalysis(BaseModel):
    key_trends: List[str] = Field(description="List of identified market trends")
    market_size: str = Field(description="Estimated market size")
    competitors: List[str] = Field(description="Major competitors in the space")


# Define flow state
class MarketResearchState(BaseModel):
    product: str = ""
    analysis: MarketAnalysis | None = None


# Create a flow class
class MarketResearchFlow(Flow[MarketResearchState]):
    @start()
    def initialize_research(self) -> Dict[str, Any]:
        print(f"Starting market research for {self.state.product}")
        return {"product": self.state.product}

    @listen(initialize_research)
    async def analyze_market(self) -> Dict[str, Any]:
        # Create an Agent for market research
        analyst = Agent(
            role="Market Research Analyst",
            goal=f"Analyze the market for {self.state.product}",
            backstory="You are an experienced market analyst with expertise in "
            "identifying market trends and opportunities.",
            tools=[SerperDevTool()],
            verbose=True,
        )

        # Define the research query
        query = f"""
        Research the market for {self.state.product}. Include:
        1. Key market trends
        2. Market size
        3. Major competitors

        Format your response according to the specified structure.
        """

        # Execute the analysis with structured output format
        result = await analyst.kickoff_async(query, response_format=MarketAnalysis)
        if result.pydantic:
            print("result", result.pydantic)
        else:
            print("result", result)

        # Return the analysis to update the state
        return {"analysis": result.pydantic}

    @listen(analyze_market)
    def present_results(self, analysis) -> None:
        print("\nMarket Analysis Results")
        print("=====================")

        if isinstance(analysis, dict):
            # If we got a dict with 'analysis' key, extract the actual analysis object
            market_analysis = analysis.get("analysis")
        else:
            market_analysis = analysis

        if market_analysis and isinstance(market_analysis, MarketAnalysis):
            print("\nKey Market Trends:")
            for trend in market_analysis.key_trends:
                print(f"- {trend}")

            print(f"\nMarket Size: {market_analysis.market_size}")

            print("\nMajor Competitors:")
            for competitor in market_analysis.competitors:
                print(f"- {competitor}")
        else:
            print("No structured analysis data available.")
            print("Raw analysis:", analysis)


# Usage example
async def run_flow():
    flow = MarketResearchFlow()
    flow.plot("MarketResearchFlowPlot")
    result = await flow.kickoff_async(inputs={"product": "AI-powered chatbots"})
    return result


# Run the flow
if __name__ == "__main__":
    asyncio.run(run_flow())

This example demonstrates several key features of using Agents in flows:

Structured Output: Using Pydantic models to define the expected output format (MarketAnalysis) ensures type safety and structured data throughout the flow.
State Management: The flow state (MarketResearchState) maintains context between steps and stores both inputs and outputs.
Tool Integration: Agents can use tools (like WebsiteSearchTool) to enhance their capabilities.

Adding Crews to Flows

Creating a flow with multiple crews in CrewAI is straightforward. You can generate a new CrewAI project that includes all the scaffolding needed to create a flow with multiple crews by running the following command:

crewai create flow name_of_flow

This command will generate a new CrewAI project with the necessary folder structure. The generated project includes a prebuilt crew called poem_crew that is already working. The starter embedded crew uses the classic Python/YAML layout; new standalone crews created with crewai create crew use the JSON-first layout.

Folder Structure

After running the crewai create flow name_of_flow command, you will see a folder structure similar to the following:

Directory/File	Description
`name_of_flow/`	Root directory for the flow.
├── `crews/`	Contains directories for specific crews.
│ └── `poem_crew/`	Directory for the “poem_crew” with its configurations and scripts.
│ ├── `config/`	Configuration files directory for the “poem_crew”.
│ │ ├── `agents.yaml`	YAML file defining the agents for “poem_crew”.
│ │ └── `tasks.yaml`	YAML file defining the tasks for “poem_crew”.
│ ├── `poem_crew.py`	Script for “poem_crew” functionality.
├── `tools/`	Directory for additional tools used in the flow.
│ └── `custom_tool.py`	Custom tool implementation.
├── `main.py`	Main script for running the flow.
├── `README.md`	Project description and instructions.
├── `pyproject.toml`	Configuration file for project dependencies and settings.
└── `.gitignore`	Specifies files and directories to ignore in version control.

Building Your Crews

In the crews folder, you can define multiple crews. The generated poem_crew uses the classic embedded-crew structure:

config/agents.yaml: Defines the agents for the crew.
config/tasks.yaml: Defines the tasks for the crew.
poem_crew.py: Contains the crew definition, including agents, tasks, and the crew itself.

You can copy, paste, and edit the poem_crew to create other classic embedded crews. For JSON-first embedded crews, use a folder with crew.jsonc and agents/*.jsonc instead:

crews/
└── research_crew/
    ├── agents/
    │   └── researcher.jsonc
    └── crew.jsonc

Then load it from a Flow step:

from pathlib import Path
from crewai.project import load_crew

crew, default_inputs = load_crew(
    Path(__file__).parent / "crews" / "research_crew" / "crew.jsonc"
)
result = crew.kickoff(inputs={**default_inputs, "topic": "AI Agents"})

Connecting Crews in `main.py`

The main.py file is where you create your flow and connect the crews together. You can define your flow by using the Flow class and the decorators @start and @listen to specify the flow of execution. Here’s an example of how you can connect the poem_crew in the main.py file:

Code

#!/usr/bin/env python
from random import randint

from pydantic import BaseModel
from crewai.flow.flow import Flow, listen, start
from .crews.poem_crew.poem_crew import PoemCrew

class PoemState(BaseModel):
    sentence_count: int = 1
    poem: str = ""

class PoemFlow(Flow[PoemState]):

    @start()
    def generate_sentence_count(self):
        print("Generating sentence count")
        self.state.sentence_count = randint(1, 5)

    @listen(generate_sentence_count)
    def generate_poem(self):
        print("Generating poem")
        result = PoemCrew().crew().kickoff(inputs={"sentence_count": self.state.sentence_count})

        print("Poem generated", result.raw)
        self.state.poem = result.raw

    @listen(generate_poem)
    def save_poem(self):
        print("Saving poem")
        with open("poem.txt", "w") as f:
            f.write(self.state.poem)

def kickoff():
    poem_flow = PoemFlow()
    poem_flow.kickoff()


def plot():
    poem_flow = PoemFlow()
    poem_flow.plot("PoemFlowPlot")

if __name__ == "__main__":
    kickoff()
    plot()

In this example, the PoemFlow class defines a flow that generates a sentence count, uses the PoemCrew to generate a poem, and then saves the poem to a file. The flow is kicked off by calling the kickoff() method. The PoemFlowPlot will be generated by plot() method.

Running the Flow

(Optional) Before running the flow, you can install the dependencies by running:

crewai install

Once all of the dependencies are installed, you need to activate the virtual environment by running:

source .venv/bin/activate

After activating the virtual environment, you can run the flow with the CrewAI CLI:

crewai run

You can also run the project script directly:

uv run kickoff

The flow will execute, and you should see the output in the console.

Plot Flows

Visualizing your AI workflows can provide valuable insights into the structure and execution paths of your flows. CrewAI offers a powerful visualization tool that allows you to generate interactive plots of your flows, making it easier to understand and optimize your AI workflows.

What are Plots?

Plots in CrewAI are graphical representations of your AI workflows. They display the various tasks, their connections, and the flow of data between them. This visualization helps in understanding the sequence of operations, identifying bottlenecks, and ensuring that the workflow logic aligns with your expectations.

How to Generate a Plot

CrewAI provides two convenient methods to generate plots of your flows:

Option 1: Using the `plot()` Method

If you are working directly with a flow instance, you can generate a plot by calling the plot() method on your flow object. This method will create an HTML file containing the interactive plot of your flow.

Code

# Assuming you have a flow instance
flow.plot("my_flow_plot")

This will generate a file named my_flow_plot.html in your current directory. You can open this file in a web browser to view the interactive plot.

Option 2: Using the Command Line

If you are working within a structured CrewAI project, you can generate a plot using the command line. This is particularly useful for larger projects where you want to visualize the entire flow setup.

crewai flow plot

This command will generate an HTML file with the plot of your flow, similar to the plot() method. The file will be saved in your project directory, and you can open it in a web browser to explore the flow.

Understanding the Plot

The generated plot will display nodes representing the tasks in your flow, with directed edges indicating the flow of execution. The plot is interactive, allowing you to zoom in and out, and hover over nodes to see additional details. By visualizing your flows, you can gain a clearer understanding of the workflow’s structure, making it easier to debug, optimize, and communicate your AI processes to others.

Conclusion

Plotting your flows is a powerful feature of CrewAI that enhances your ability to design and manage complex AI workflows. Whether you choose to use the plot() method or the command line, generating plots will provide you with a visual representation of your workflows, aiding in both development and presentation.

Next Steps

If you’re interested in exploring additional examples of flows, we have a variety of recommendations in our examples repository. Here are four specific flow examples, each showcasing unique use cases to help you match your current problem type to a specific example:

Email Auto Responder Flow: This example demonstrates an infinite loop where a background job continually runs to automate email responses. It’s a great use case for tasks that need to be performed repeatedly without manual intervention. View Example
Lead Score Flow: This flow showcases adding human-in-the-loop feedback and handling different conditional branches using the router. It’s an excellent example of how to incorporate dynamic decision-making and human oversight into your workflows. View Example
Write a Book Flow: This example excels at chaining multiple crews together, where the output of one crew is used by another. Specifically, one crew outlines an entire book, and another crew generates chapters based on the outline. Eventually, everything is connected to produce a complete book. This flow is perfect for complex, multi-step processes that require coordination between different tasks. View Example
Meeting Assistant Flow: This flow demonstrates how to broadcast one event to trigger multiple follow-up actions. For instance, after a meeting is completed, the flow can update a Trello board, send a Slack message, and save the results. It’s a great example of handling multiple outcomes from a single event, making it ideal for comprehensive task management and notification systems. View Example

By exploring these examples, you can gain insights into how to leverage CrewAI Flows for various use cases, from automating repetitive tasks to managing complex, multi-step processes with dynamic decision-making and human feedback. Also, check out our YouTube video on how to use flows in CrewAI below!

Running Flows

There are two ways to run a flow:

Using the Flow API

You can run a flow programmatically by creating an instance of your flow class and calling the kickoff() method:

flow = ExampleFlow()
result = flow.kickoff()

Streaming Flow Execution

For real-time visibility into flow execution, you can enable streaming to receive output as it’s generated:

class StreamingFlow(Flow):
    stream = True  # Enable streaming

    @start()
    def research(self):
        # Your flow implementation
        pass

# Iterate over streaming output
flow = StreamingFlow()
streaming = flow.kickoff()
for chunk in streaming:
    print(chunk.content, end="", flush=True)

# Access final result
result = streaming.result

Learn more about streaming in the Streaming Flow Execution guide.

Memory in Flows

Every Flow automatically has access to CrewAI’s unified Memory system. You can store, recall, and extract memories directly inside any flow method using three built-in convenience methods.

Built-in Methods

Method	Description
`self.remember(content, **kwargs)`	Store content in memory. Accepts optional `scope`, `categories`, `metadata`, `importance`.
`self.recall(query, **kwargs)`	Retrieve relevant memories. Accepts optional `scope`, `categories`, `limit`, `depth`.
`self.extract_memories(content)`	Break raw text into discrete, self-contained memory statements.

A default Memory() instance is created automatically when the Flow initializes. You can also pass a custom one:

from crewai.flow.flow import Flow
from crewai import Memory

custom_memory = Memory(
    recency_weight=0.5,
    recency_half_life_days=7,
    embedder={"provider": "ollama", "config": {"model_name": "mxbai-embed-large"}},
)

flow = MyFlow(memory=custom_memory)

Example: Research and Analyze Flow

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


class ResearchAnalysisFlow(Flow):
    @start()
    def gather_data(self):
        # Simulate research findings
        findings = (
            "PostgreSQL handles 10k concurrent connections with connection pooling. "
            "MySQL caps at around 5k. MongoDB scales horizontally but adds complexity."
        )

        # Extract atomic facts and remember each one
        memories = self.extract_memories(findings)
        for mem in memories:
            self.remember(mem, scope="/research/databases")

        return findings

    @listen(gather_data)
    def analyze(self, raw_findings):
        # Recall relevant past research (from this run or previous runs)
        past = self.recall("database performance and scaling", limit=10, depth="shallow")

        context_lines = [f"- {m.record.content}" for m in past]
        context = "\n".join(context_lines) if context_lines else "No prior context."

        return {
            "new_findings": raw_findings,
            "prior_context": context,
            "total_memories": len(past),
        }


flow = ResearchAnalysisFlow()
result = flow.kickoff()
print(result)

Because memory persists across runs (backed by LanceDB on disk), the analyze step will recall findings from previous executions too — enabling flows that learn and accumulate knowledge over time. See the Memory documentation for details on scopes, slices, composite scoring, embedder configuration, and more.

Using the CLI

Starting from version 0.103.0, you can run flows using the crewai run command:

crewai run

This command automatically detects if your project is a flow (based on the type = "flow" setting in your pyproject.toml) and runs it accordingly. This is the recommended way to run flows from the command line. The legacy crewai flow kickoff command is deprecated. Use crewai run for both crews and flows.

​Overview

​Getting Started

​@start()

​@listen()

​Usage

​Flow Output

​Retrieving the Final Output

​Accessing and Updating State

​Flow Usage Metrics

​Flow State Management

​Unstructured State Management

​Structured State Management

​Choosing Between Unstructured and Structured State Management

​Flow Persistence

​Class-Level Persistence

​Method-Level Persistence

​Forking Persisted State

​How It Works

​Important Considerations

​Technical Advantages

​Flow Control

​Conditional Logic: or

​Conditional Logic: and

​Router

​Human in the Loop (human feedback)

​Adding Agents to Flows

​Adding Crews to Flows

​Folder Structure

​Building Your Crews

​Connecting Crews in main.py

​Running the Flow

​Plot Flows

​What are Plots?

​How to Generate a Plot

​Option 1: Using the plot() Method

​Option 2: Using the Command Line

​Understanding the Plot

​Conclusion

​Next Steps

​Running Flows

​Using the Flow API

​Streaming Flow Execution

​Memory in Flows

​Built-in Methods

​Example: Research and Analyze Flow

​Using the CLI

Overview

Getting Started

@start()

@listen()

Usage

Flow Output

Retrieving the Final Output

Accessing and Updating State

Flow Usage Metrics

Flow State Management

Unstructured State Management

Structured State Management

Choosing Between Unstructured and Structured State Management

Flow Persistence

Class-Level Persistence

Method-Level Persistence

Forking Persisted State

How It Works

Important Considerations

Technical Advantages

Flow Control

Conditional Logic: `or`

Conditional Logic: `and`

Router

Human in the Loop (human feedback)

Adding Agents to Flows

Adding Crews to Flows

Folder Structure

Building Your Crews

Connecting Crews in `main.py`

Running the Flow

Plot Flows

What are Plots?

How to Generate a Plot

Option 1: Using the `plot()` Method

Option 2: Using the Command Line

Understanding the Plot

Conclusion

Next Steps

Running Flows

Using the Flow API

Streaming Flow Execution

Memory in Flows

Built-in Methods

Example: Research and Analyze Flow

Using the CLI