Workflows are deterministic, stateful, multi-agent pipelines that power many of our production use cases. They are incredibly powerful and offer the following benefits:

  • Control and Flexibility: You have full control over the multi-agent process, how the input is processed, which agents are used and in what order.
  • Built-in Memory: You can store state and cache results in a database at any time, meaning your agents can re-use results from previous steps.
  • Defined as a python class: You do not need to learn a new framework, its just python.

How to build a workflow:

  1. Define your workflow as a class by inheriting from the Workflow class
  2. Add one or more agents to the workflow
  3. Implement your logic in the run() method
  4. Cache results in the session_state as needed
  5. Run the workflow using the .run() method

​
Example: Blog Post Generator

Let’s create a blog post generator that can search the web, read the top links and write a blog post for us. We’ll cache intermediate results in the database to improve performance.

​
Create the Workflow

Create a file blog_post_generator.py

blog_post_generator.py
import json
from typing import Optional, Iterator

from pydantic import BaseModel, Field

from phi.agent import Agent
from phi.workflow import Workflow, RunResponse, RunEvent
from phi.storage.workflow.sqlite import SqlWorkflowStorage
from phi.tools.duckduckgo import DuckDuckGo
from phi.utils.pprint import pprint_run_response
from phi.utils.log import logger


class NewsArticle(BaseModel):
    title: str = Field(..., description="Title of the article.")
    url: str = Field(..., description="Link to the article.")
    summary: Optional[str] = Field(..., description="Summary of the article if available.")


class SearchResults(BaseModel):
    articles: list[NewsArticle]


class BlogPostGenerator(Workflow):
    searcher: Agent = Agent(
        tools=[DuckDuckGo()],
        instructions=["Given a topic, search for 20 articles and return the 5 most relevant articles."],
        response_model=SearchResults,
    )

    writer: Agent = Agent(
        instructions=[
            "You will be provided with a topic and a list of top articles on that topic.",
            "Carefully read each article and generate a New York Times worthy blog post on that topic.",
            "Break the blog post into sections and provide key takeaways at the end.",
            "Make sure the title is catchy and engaging.",
            "Always provide sources, do not make up information or sources.",
        ],
    )

    def run(self, topic: str, use_cache: bool = True) -> Iterator[RunResponse]:
        logger.info(f"Generating a blog post on: {topic}")

        # Use the cached blog post if use_cache is True
        if use_cache and "blog_posts" in self.session_state:
            logger.info("Checking if cached blog post exists")
            for cached_blog_post in self.session_state["blog_posts"]:
                if cached_blog_post["topic"] == topic:
                    logger.info("Found cached blog post")
                    yield RunResponse(
                        run_id=self.run_id,
                        event=RunEvent.workflow_completed,
                        content=cached_blog_post["blog_post"],
                    )
                    return

        # Step 1: Search the web for articles on the topic
        num_tries = 0
        search_results: Optional[SearchResults] = None
        # Run until we get a valid search results
        while search_results is None and num_tries < 3:
            try:
                num_tries += 1
                searcher_response: RunResponse = self.searcher.run(topic)
                if (
                    searcher_response
                    and searcher_response.content
                    and isinstance(searcher_response.content, SearchResults)
                ):
                    logger.info(f"Searcher found {len(searcher_response.content.articles)} articles.")
                    search_results = searcher_response.content
                else:
                    logger.warning("Searcher response invalid, trying again...")
            except Exception as e:
                logger.warning(f"Error running searcher: {e}")

        # If no search_results are found for the topic, end the workflow
        if search_results is None or len(search_results.articles) == 0:
            yield RunResponse(
                run_id=self.run_id,
                event=RunEvent.workflow_completed,
                content=f"Sorry, could not find any articles on the topic: {topic}",
            )
            return

        # Step 2: Write a blog post
        logger.info("Writing blog post")
        # Prepare the input for the writer
        writer_input = {
            "topic": topic,
            "articles": [v.model_dump() for v in search_results.articles],
        }
        # Run the writer and yield the response
        yield from self.writer.run(json.dumps(writer_input, indent=4), stream=True)

        # Save the blog post in the session state for future runs
        if "blog_posts" not in self.session_state:
            self.session_state["blog_posts"] = []
        self.session_state["blog_posts"].append({"topic": topic, "blog_post": self.writer.run_response.content})


# The topic to generate a blog post on
topic = "US Elections 2024"

# Create the workflow
generate_blog_post = BlogPostGenerator(
    session_id=f"generate-blog-post-on-{topic}",
    storage=SqlWorkflowStorage(
        table_name="generate_blog_post_workflows",
        db_file="tmp/workflows.db",
    ),
)

# Run workflow
blog_post: Iterator[RunResponse] = generate_blog_post.run(topic=topic, use_cache=True)

# Print the response
pprint_run_response(blog_post, markdown=True)

​
Run the workflow

Install libraries

pip install phidata openai duckduckgo-search sqlalchemy phidata

Run the workflow

python blog_post_generator.py

Now the results are cached in the database and can be re-used for future runs. Run the workflow again to view the cached results.

python blog_post_generator.py

Workflows are deterministic, stateful, multi-agent pipelines that power many of our production use cases. They are incredibly powerful and offer the following benefits:

  • Control and Flexibility: You have full control over the multi-agent process, how the input is processed, which agents are used and in what order.
  • Built-in Memory: You can store state and cache results in a database at any time, meaning your agents can re-use results from previous steps.
  • Defined as a python class: You do not need to learn a new framework, its just python.

How to build a workflow:

  1. Define your workflow as a class by inheriting from the Workflow class
  2. Add one or more agents to the workflow
  3. Implement your logic in the run() method
  4. Cache results in the session_state as needed
  5. Run the workflow using the .run() method

​
Example: Blog Post Generator

Let’s create a blog post generator that can search the web, read the top links and write a blog post for us. We’ll cache intermediate results in the database to improve performance.

​
Create the Workflow

Create a file blog_post_generator.py

blog_post_generator.py
import json
from typing import Optional, Iterator

from pydantic import BaseModel, Field

from phi.agent import Agent
from phi.workflow import Workflow, RunResponse, RunEvent
from phi.storage.workflow.sqlite import SqlWorkflowStorage
from phi.tools.duckduckgo import DuckDuckGo
from phi.utils.pprint import pprint_run_response
from phi.utils.log import logger


class NewsArticle(BaseModel):
    title: str = Field(..., description="Title of the article.")
    url: str = Field(..., description="Link to the article.")
    summary: Optional[str] = Field(..., description="Summary of the article if available.")


class SearchResults(BaseModel):
    articles: list[NewsArticle]


class BlogPostGenerator(Workflow):
    searcher: Agent = Agent(
        tools=[DuckDuckGo()],
        instructions=["Given a topic, search for 20 articles and return the 5 most relevant articles."],
        response_model=SearchResults,
    )

    writer: Agent = Agent(
        instructions=[
            "You will be provided with a topic and a list of top articles on that topic.",
            "Carefully read each article and generate a New York Times worthy blog post on that topic.",
            "Break the blog post into sections and provide key takeaways at the end.",
            "Make sure the title is catchy and engaging.",
            "Always provide sources, do not make up information or sources.",
        ],
    )

    def run(self, topic: str, use_cache: bool = True) -> Iterator[RunResponse]:
        logger.info(f"Generating a blog post on: {topic}")

        # Use the cached blog post if use_cache is True
        if use_cache and "blog_posts" in self.session_state:
            logger.info("Checking if cached blog post exists")
            for cached_blog_post in self.session_state["blog_posts"]:
                if cached_blog_post["topic"] == topic:
                    logger.info("Found cached blog post")
                    yield RunResponse(
                        run_id=self.run_id,
                        event=RunEvent.workflow_completed,
                        content=cached_blog_post["blog_post"],
                    )
                    return

        # Step 1: Search the web for articles on the topic
        num_tries = 0
        search_results: Optional[SearchResults] = None
        # Run until we get a valid search results
        while search_results is None and num_tries < 3:
            try:
                num_tries += 1
                searcher_response: RunResponse = self.searcher.run(topic)
                if (
                    searcher_response
                    and searcher_response.content
                    and isinstance(searcher_response.content, SearchResults)
                ):
                    logger.info(f"Searcher found {len(searcher_response.content.articles)} articles.")
                    search_results = searcher_response.content
                else:
                    logger.warning("Searcher response invalid, trying again...")
            except Exception as e:
                logger.warning(f"Error running searcher: {e}")

        # If no search_results are found for the topic, end the workflow
        if search_results is None or len(search_results.articles) == 0:
            yield RunResponse(
                run_id=self.run_id,
                event=RunEvent.workflow_completed,
                content=f"Sorry, could not find any articles on the topic: {topic}",
            )
            return

        # Step 2: Write a blog post
        logger.info("Writing blog post")
        # Prepare the input for the writer
        writer_input = {
            "topic": topic,
            "articles": [v.model_dump() for v in search_results.articles],
        }
        # Run the writer and yield the response
        yield from self.writer.run(json.dumps(writer_input, indent=4), stream=True)

        # Save the blog post in the session state for future runs
        if "blog_posts" not in self.session_state:
            self.session_state["blog_posts"] = []
        self.session_state["blog_posts"].append({"topic": topic, "blog_post": self.writer.run_response.content})


# The topic to generate a blog post on
topic = "US Elections 2024"

# Create the workflow
generate_blog_post = BlogPostGenerator(
    session_id=f"generate-blog-post-on-{topic}",
    storage=SqlWorkflowStorage(
        table_name="generate_blog_post_workflows",
        db_file="tmp/workflows.db",
    ),
)

# Run workflow
blog_post: Iterator[RunResponse] = generate_blog_post.run(topic=topic, use_cache=True)

# Print the response
pprint_run_response(blog_post, markdown=True)

​
Run the workflow

Install libraries

pip install phidata openai duckduckgo-search sqlalchemy phidata

Run the workflow

python blog_post_generator.py

Now the results are cached in the database and can be re-used for future runs. Run the workflow again to view the cached results.

python blog_post_generator.py