Building a Multi-Agent AI System from Scratch (No Frameworks)

#ai #python #agents #tutorial

You dont need CrewAI or AutoGen. Heres how to build a multi-agent pipeline in pure Python.

The Concept

5 AI agents collaborate in a pipeline:

Researcher → Writer → Editor → Reviewer → Publisher

Each agent has a role, a system prompt, and passes output to the next.

Implementation

class Agent:
    def __init__(self, name, role, system_prompt):
        self.name = name
        self.role = role
        self.system_prompt = system_prompt

    def process(self, input_text):
        """Call LLM with system prompt + input"""
        import requests
        response = requests.post(
            "http://localhost:11434/api/generate",
            json={
                "model": "llama3.2",
                "prompt": f"{self.system_prompt}\n\nInput: {input_text}",
                "stream": False
            }
        )
        return response.json()["response"]


class Pipeline:
    def __init__(self, agents):
        self.agents = agents

    def run(self, topic):
        output = topic
        for agent in self.agents:
            print(f"🤖 {agent.name} working...")
            output = agent.process(output)
            print(f"   ✅ Done ({len(output)} chars)")
        return output


# Define agents
agents = [
    Agent("Researcher", "research",
          "Find 5 key facts about the topic. Be concise."),
    Agent("Writer", "write",
          "Write a 500-word article from the research."),
    Agent("Editor", "edit",
          "Improve clarity, fix grammar, make it engaging."),
    Agent("Reviewer", "review",
          "Score 1-10. List what works and what needs fixing."),
    Agent("Publisher", "publish",
          "Format as markdown with title, headings, and CTA."),
]

# Run
pipeline = Pipeline(agents)
result = pipeline.run("The Future of Local AI")
print(result)

Why No Framework?

Simpler: 50 lines vs 500 lines with CrewAI
Debuggable: You can see exactly whats happening
Flexible: Add/remove agents easily
No dependencies: Just requests library
Works with any LLM: Ollama, OpenAI, Groq, anything

Pro Tips

Different models per agent: Use codellama for code review, llama3.2 for writing
Temperature: Lower for Researcher (factual), higher for Writer (creative)
Fallback: If one model fails, try another
Logging: Save each agents output for debugging