M10-A — When One Agent Is Not Enough

How to read this note

One smart person can do many jobs — but slowly, and not all at once. A film crew has director, camera, sound, editor — each expert. Multi-agent AI works the same way.

🎬 The film crew analogy

Film role	Agent role
Director	Manager agent — decides order of work
Script writer	Research agent — gathers facts
Camera	Tool agent — fetches web/files
Editor	Writer agent — polishes final output
Producer	Human you — approves release

One actor cannot shoot a whole movie alone. One general AI agent struggles on big projects alone too.

🤔 When do you need multiple agents?

One agent enough	Multiple agents better
"Summarise this page"	"Research market → write report → make slides → email client"
Single skill task	Needs research + writing + checking
Short output	Long pipeline with quality gates
You guide every step	Different roles need different prompts

👥 Collaboration patterns

Pattern 1 — Sequential handoff

Research Agent  →  passes notes  →  Writer Agent  →  passes draft  →  Critic Agent

Like an assembly line. Each agent sees only its job + previous output.

Pattern 2 — Manager + workers

        Manager Agent
       /      |      \
Research   Writer   Fact-checker
   Agent     Agent      Agent

Manager breaks goal into subtasks, assigns agents, collects results.

Pattern 3 — Debate / review

Agent A writes proposal
Agent B finds flaws
Agent A revises
Repeat until quality OK

Good for important documents — still needs human final sign-off.

⚙️ What happens in the backend

┌─────────────────────────────────────────┐
│           ORCHESTRATOR                  │
│  (CrewAI / AutoGen / custom code)       │
└─────────┬───────────┬───────────┬───────┘
          ▼           ▼           ▼
     ┌────────┐  ┌────────┐  ┌────────┐
     │Agent 1 │  │Agent 2 │  │Agent 3 │
     │ LLM +  │  │ LLM +  │  │ LLM +  │
     │ prompt │  │ prompt │  │ prompt │
     └────────┘  └────────┘  └────────┘

Orchestrator = traffic police. It decides:

Which agent runs next
What message passes between them
When to stop

Each agent may use the same ChatGPT/Gemini model with a different role prompt — or different models for cost/speed.

The real insight

Multi-agent is not "more AI brains." It is clearer job descriptions — each agent does one thing well.

📋 Example — Launch a product poster campaign

Agent	Role prompt summary	Output
Market	"List 3 audience pain points for this product"	Bullet list
Copy	"Write poster headline from these pain points"	Headline options
Design brief	"Turn winning headline into image AI prompt"	Image prompt
QA	"Check claims against product PDF — flag lies"	Pass / fail

Human picks headline and approves poster.

⚠️ Common mistakes

Mistake	Fix
10 agents for a 1-agent job	Start with 2–3 roles max
Agents argue forever	Set max rounds (example: 2 revisions)
No shared memory format	Standard handoff template (Module 5)
No human checkpoint	Approve before anything customer-facing

Next: M10-B — Task Distribution and Orchestration
Module 11: M11-A — CrewAI How It Works — build this in code/no-code tools