Skip to main content

Multi-Agent Collaboration: AutoGen & CrewAI

When tasks exceed the capability of a single prompt or agent loop, Multi-Agent Frameworks coordinate specialized teams of agents with complementary roles, skills, and tools.

Multi-Agent Advantage

Assigning specific personas (e.g. Senior Software Architect, Security Auditor, Code Writer) reduces prompt dilution and improves complex problem solving.

1. Microsoft AutoGen​

AutoGen enables conversational multi-agent systems where agents communicate via message streams.

  • AssistantAgent: Performs tasks, writes code, and calls tools.
  • UserProxyAgent: Represents the human user, executes generated code in a sandbox, and provides feedback.
  • GroupChatManager: Coordinates multi-agent conversations between multiple specialized agents.
from autogen import AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager

coder = AssistantAgent(name="Coder", llm_config=llm_config)
reviewer = AssistantAgent(name="Code_Reviewer", llm_config=llm_config)
user_proxy = UserProxyAgent(name="User", code_execution_config={"work_dir": "workspace"})

group_chat = GroupChat(agents=[user_proxy, coder, reviewer], messages=[], max_round=12)
manager = GroupChatManager(groupchat=group_chat, llm_config=llm_config)

user_proxy.initiate_chat(manager, message="Build a FastAPI microservice for matrix multiplication.")

2. CrewAI​

CrewAI enforces role-based, structured team hierarchies:

  • Agents: Configured with explicit role, goal, and backstory.
  • Tasks: Discrete units of work assigned to specific agents with expected outputs.
  • Crew: The orchestrator running tasks sequentially or hierarchically.
from crewai import Agent, Task, Crew, Process

researcher = Agent(
    role="AI Research Analyst",
    goal="Discover cutting edge papers on local quantization",
    backstory="You are a veteran AI researcher tracking hardware efficiency trends."
)

task1 = Task(
    description="Analyze recent GGUF 4-bit quantization benchmarks for mobile.",
    agent=researcher,
    expected_output="A bulleted summary of memory vs accuracy tradeoffs."
)

crew = Crew(
    agents=[researcher],
    tasks=[task1],
    process=Process.sequential
)

result = crew.kickoff()