M11-B — AutoGen Introduction

How to read this note

AutoGen is another multi-agent framework (from Microsoft research). If CrewAI feels like assigning job tickets, AutoGen often feels like agents talking in a group chat until the job is done.

💬 Core idea — agents in conversation

User: "Build a study plan for Module 7 exam"

Planner Agent:  "First list all RAG topics from syllabus"
Tutor Agent:    "Here are 6 topics with difficulty order"
Critic Agent:   "Add more time for vector databases — students struggle"
Planner Agent:  "Updated 7-day plan attached"

Agents message each other. The conversation is the workflow.

🆚 CrewAI vs AutoGen (simple table)

	CrewAI	AutoGen
Style	Task list + roles	Chat between agents
Best for	Clear pipelines	Brainstorming, coding help
Process	Sequential / hierarchical tasks	Group chat rounds
Learning curve	Medium	Medium–high
Feel	Project manager office	WhatsApp group of experts

Both use LLM APIs in the backend. Both can use tools.

You do not need to master both — brochure says "CrewAI / AutoGen basics" — pick one for your capstone.

⚙️ AutoGen backend (what runs)

┌─────────────────────────────────────┐
│         Group Chat Manager          │
│  (decides who speaks next)          │
└───────────┬─────────────────────────┘
            │
     ┌──────┼──────┐
     ▼      ▼      ▼
  Agent A  Agent B  User proxy (you)
     │      │      │
     └──────┴──────┘
            │
     Shared chat history = memory
            │
            ▼
     LLM API calls each turn

Each "message" in the group → one LLM call with full chat history + that agent's system prompt.

Stop when:

Max rounds reached
Agent sends TERMINATE signal
Human says stop

🧩 Key AutoGen pieces

Piece	Job
AssistantAgent	AI worker (researcher, coder, writer)
UserProxyAgent	Represents you — can run code on your machine
GroupChat	Shared room for agents
GroupChatManager	Picks next speaker

UserProxyAgent — special power

Can execute Python code the assistant writes (if you allow it).

Coder Agent: "Here's Python to plot your WhatsApp stats"
User Proxy:  runs code locally → shows chart

Powerful for data projects. Dangerous if unrestricted — only enable in trusted exercises.

🎯 When to choose AutoGen

Choose AutoGen when…	Choose CrewAI when…
Agents need to debate or refine	Steps are fixed order
Coding + execution loop	Marketing / research pipelines
Exploratory work	Repeatable business process

📋 Same project in both frameworks (concept)

Goal: Weekly quiz from class notes

CrewAI version:

Task 1: Extract topics → Task 2: Write 10 MCQs → Task 3: Verify answers

AutoGen version:

Tutor + Fact-checker chat until both agree quiz is accurate

Same outcome. Different conversation shape.

⚠️ Safety reminders

Limit code execution sandbox
Set max_round on group chats
Never put passwords in agent chat
Human reviews output before sharing with class

✅ Module 11 summary

Brochure Agent Frameworks (Intro):

Tool	You learned
CrewAI	Agents + Tasks + Crew + sequential backend loop
AutoGen	Group chat agents + manager + optional code run

Next phase (Module 12–13): connect agents to n8n, websites, and deployment.

Previous: M11-A — CrewAI How It Works
Automation: N8N install · Open claw
Capstone: See Homepage — choose Student / Professional / Entrepreneur track