Module 11: Multi-Agent Systems | Generative AI & LLM

Duration: 4 hours | Difficulty: Advanced | Prerequisites: Modules 08–10

Learning Objectives

By the end of this module, you will be able to:

Design multi-agent systems with supervisor, swarm, and peer-to-peer patterns
Build collaborative agent teams using CrewAI
Implement conversational agent groups with AutoGen
Handle inter-agent communication, task delegation, and conflict resolution
Debug and monitor multi-agent workflows

1. Why Multi-Agent?

A single agent runs out of focus and context on big projects. Multi-agent setups give you roles: researcher, writer, critic, each with narrower prompts and tools.

Analogy: building a house—electrician vs plumber vs framer. Same job site, different expertise.

Single Agent:         One LLM does everything (overloaded)

Multi-Agent:          ┌─ Researcher ─────┐
                      │                   ▼
         Supervisor ──┼─ Writer ──────▶ Editor ──▶ Final Output
                      │                   ▲
                      └─ Fact-Checker ───┘

Multi-Agent Patterns

Pattern	Structure	Best for
Supervisor	Coordinator delegates	Predictable workflows
Swarm	Emergent coordination	Brainstormy exploration
Peer-to-Peer	Agents talk laterally	Debate / consensus
Hierarchical	Layers of managers	Big org simulations

Try This! For your next project, list three distinct responsibilities—if they argue in one prompt, they might deserve separate agents.

2. CrewAI: Task-Oriented Agent Teams

CrewAI models Agents, Tasks, and Crews (with sequential or hierarchical process). Tasks can declare context dependencies—outputs flow like a Makefile.

CrewAI with Tools

Give agents tools the same way you would in single-agent apps—web search wrappers, file readers, SQL—then let the crew orchestration handle ordering.

Key Example: Tiny three-role crew: research → write → edit. Trim verbosity in real runs.

python

[object Object], crewai ,[object Object], Agent, Task, Crew, Process
,[object Object], langchain_openai ,[object Object], ChatOpenAI

llm = ChatOpenAI(model=,[object Object],, temperature=,[object Object],, api_key=,[object Object],)

researcher = Agent(
    role=,[object Object],,
    goal=,[object Object],,
    backstory=,[object Object],,
    llm=llm,
)
writer = Agent(
    role=,[object Object],,
    goal=,[object Object],,
    backstory=,[object Object],,
    llm=llm,
)
editor = Agent(
    role=,[object Object],,
    goal=,[object Object],,
    backstory=,[object Object],,
    llm=llm,
)

t1 = Task(
    description=,[object Object],,
    expected_output=,[object Object],,
    agent=researcher,
)
t2 = Task(
    description=,[object Object],,
    expected_output=,[object Object],,
    agent=writer,
    context=[t1],
)
t3 = Task(
    description=,[object Object],,
    expected_output=,[object Object],,
    agent=editor,
    context=[t2],
)

crew = Crew(agents=[researcher, writer, editor], tasks=[t1, t2, t3], process=Process.sequential)
,[object Object],(crew.kickoff())

3. AutoGen: Conversational Multi-Agent

AutoGen-style frameworks let agents send messages to each other in a group chat. Good for iterative critique (“coder” vs “reviewer”) and role-play simulations.

CrewAI vibe	AutoGen vibe
Task graph	Chat transcript
Great for ETL-like pipelines	Great for debate loops

Concept: Pick the framework that matches your control flow—DAG vs conversation.

4. Supervisor Pattern with LangGraph

Implement supervisor as a node that picks the next worker using structured output or function calls. Workers return to supervisor until END. This mirrors Module 10 graphs—multi-agent is mostly orchestration.

Tip	Why
Log decisions	Debug “why did it pick writer2?”
Cap handoffs	Prevent infinite ping-pong

5. Debugging Multi-Agent Systems

Symptom	Likely cause
Circular chatter	Missing stop condition
Duplicate work	Task context not wired
Drifting tone	Shared LLM temperature too high
Cost spike	Parallel agents + huge prompts

Observability: trace each agent call (LangSmith, OpenTelemetry, or plain structured logs).

Key Takeaway