Step 8: Run the Pipeline — Create a Multi-Agent System

One-Line Summary: Write the main entry point that orchestrates all three agents in sequence, execute the full pipeline, and understand the output.

Prerequisites: Steps 1-7 completed, all agents defined

The Orchestration

Our pipeline is a simple sequence: Researcher → Writer → Editor. Each agent's output becomes the next agent's input. The orchestration code is straightforward:

# main.py — Entry point for the content production pipeline
 
import sys
from agents import create_researcher, create_writer, create_editor
 
 
def run_pipeline(topic: str) -> str:
    """Run the three-agent content production pipeline."""
 
    # Step 1: Create the agents
    researcher = create_researcher()
    writer = create_writer()
    editor = create_editor()
 
    # Step 2: Research
    print(f"\nTopic: {topic}")
    print(f"{'='*60}")
 
    research_notes = researcher.run(
        f"Research the following topic thoroughly: {topic}"
    )
 
    # Step 3: Write (receives research notes as context)
    draft = writer.run(
        f"Write a comprehensive article about: {topic}\n\n"
        f"Use these research notes as your source material:\n\n"
        f"{research_notes}"
    )
 
    # Step 4: Edit and save (receives the draft)
    final = editor.run(
        f"Edit and polish this draft article about: {topic}\n\n"
        f"After editing, save the final version using the save_to_file tool.\n\n"
        f"Draft to edit:\n\n{draft}"
    )
 
    print(f"\n{'='*60}")
    print("Pipeline complete!")
    print(f"{'='*60}")
 
    return final
 
 
if __name__ == "__main__":
    if len(sys.argv) > 1:
        topic = " ".join(sys.argv[1:])
    else:
        topic = "The Future of Quantum Computing"
 
    output = run_pipeline(topic)
    print(output)

Run It

python main.py "The Rise of Multi-Agent AI Systems"

Or with a different topic:

python main.py "How Rust Is Changing Systems Programming"

The entire run takes 2-5 minutes depending on topic complexity and API response times.

Understanding the Output

Watch the terminal. You will see three distinct phases:

Phase 1 — Researcher at work:

============================================================
Agent: Researcher
============================================================
  [Researcher] Using tool: web_search
  [Researcher] Using tool: web_search
  [Researcher] Using tool: web_search
  [Researcher] Done.

The Researcher makes multiple search calls, refines queries based on results, and compiles structured notes.

Phase 2 — Writer at work:

============================================================
Agent: Writer
============================================================
  [Writer] Done.

The Writer receives the research notes and produces a full draft. No tool calls — just writing.

Phase 3 — Editor at work:

============================================================
Agent: Editor
============================================================
  [Editor] Using tool: save_to_file
  [Editor] Done.

The Editor polishes the draft and saves it to the output directory.

Check the Output

ls output/
cat output/*.md

The article should have a clear title, logical heading structure, specific facts from the research, and clean prose.

How This Compares to Using a Framework

Here is what our pipeline looks like versus CrewAI:

# Our approach — explicit and transparent
research_notes = researcher.run(f"Research: {topic}")
draft = writer.run(f"Write about: {topic}\n\nNotes:\n{research_notes}")
final = editor.run(f"Edit this draft:\n{draft}")

# CrewAI approach — declarative but opaque
crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    process=Process.sequential,
)
result = crew.kickoff()

Both produce the same result. The framework version is fewer lines, but our version is completely transparent — you can see exactly what each agent receives and produces. When something goes wrong, you know exactly where to look.

Troubleshooting

Issue	Cause	Fix
Agent loops without producing output	System prompt is ambiguous	Make instructions more explicit
Empty search results	Search query too narrow	Check the Researcher's queries in the logs
Writer ignores research	Notes not included in the message	Verify the string concatenation in `run_pipeline()`
Editor does not save	Tool not called	Check the Editor's system prompt mentions `save_to_file`
Rate limit errors	Too many API calls	Add a `time.sleep(1)` between agent runs

← The Editor | Next: Step 9 - What's Next →