Week 4 — Many Minds (Multi-Agent)

Today you will

Three real wins

🌳

Branch the thinking

Use Tree-of-Thought: explore several reasoning paths, keep the best.

🎭

Define roles

Give each agent a role, a goal, and only the tools it needs.

🤝

Hand off work

Wire a crew in CrewAI/LangGraph so output flows agent → agent.

Tree of thought: a GOAL branches into scored ideas; the highest-scoring branch leads to BEST — Tree-of-thought: branch into ideas, **score** them, keep the best.

Two patterns today: TREE-OF-THOUGHT = one agent explores multiple solution branches and picks the best. MULTI-AGENT = several agents with distinct roles pass work down a pipeline.

The Arc · check each off as you go in Colab (NB4)

Your step-by-step for today

One-shot vs. branch. Solve a tricky puzzle with a single prompt, then with 3 branches + a chooser. Compare.
Meet the roles. Flip the role cards below; write the goal + tools for each of your three agents.
Build the Researcher. An agent whose only job is to gather facts (it gets the search + RAG tools).
Build the Writer. An agent that turns the Researcher’s notes into a clear draft (no tools — just writing).
Wire the hand-off. Researcher → Writer → Critic, in CrewAI or a LangGraph. Run it end to end.
Route the tasks. Play the routing game below, then screenshot your crew’s final output.

The Crew

Three roles — tap a card to flip it

A good crew is like a newsroom: everyone has one job and hands work to the next person. Flip each card to see the role’s goal and the tools it should (and shouldn’t) get.

Multi-agent crew: RESEARCHER hands off to WRITER hands off to CRITIC, producing a RESULT — A crew hands work down the line: **researcher → writer → critic**.

Think like the orchestrator

Route each task to the right agent

In a multi-agent workflow, the orchestrator’s job is to send each sub-task to the specialist built for it. Which crew member should handle this?

0 / 4

Loading…

Checkpoint

Did the idea land?

Why does splitting a job across role-based agents often beat one mega-prompt?

Law #4: many small agents beat one giant prompt. Focus and clear hand-offs — not raw agent count — are what help. (More agents can also cost more, so keep the crew as small as it can be.)

Tree-of-Thought helps most when…

Exploring branches costs more time/tokens, so it pays off on hard, multi-path problems — planning, puzzles, strategy — not simple lookups.

Lock it in

Law, badge, and this week’s mission

Law #4 · Week 4

Many small agents beat one giant prompt.

Split the job, define the roles, let them hand work off. Specialists coordinate; generalists drown.

🎭

Badge earned: Orchestrator

You ran a crew of role-based agents that handed work off and produced one result.

🎯 Mission before Week 5 — start your capstone

Pick your capstone: a real problem an agent (or crew) could help with. Write the goal in one sentence.
Decide: solo agent or crew? List the roles and the tools each needs. Bring this to Week 5.

Next: Guardrails & Ship It → Week 5