Stop 15 · AI Agents & OrchestrationTopics · 5/9

Multi-Agent Systems (AutoGen, CrewAI)

In short

One agent with 30 tools gets confused. Split the work into a researcher, a writer and a reviewer, each with a small job, and quality often goes up.

Architectures where multiple specialized agents collaborate (e.g. a researcher agent, a writer agent, a critic agent) on a task too complex for a single agent to handle well alone.

The blob with a clipboard coordinates three small helpers, one painting a poster, one carrying boxes, one setting up a microphone.

Think of it like this

The college fest committee

Nobody runs a fest alone. One team gets sponsors, one designs posters, one handles stage. The fest head checks in, passes notes between teams and signs off. Multi-agent systems are the same: several focused agents, one coordinator.

Watch it happen

A small crew writes a blog post

Manager

Researcher

Writer

Reviewer

find 3 facts on UPI

1/6The manager gives the researcher a narrow task with a clear expected output.

The big ideas

What you need to know

One role, one job

Give each agent a role, a goal and only the tools it needs. Narrow agents make fewer weird choices.

Who talks to whom

Sequential (A then B), manager-led (a boss assigns tasks) or group chat. Pick the simplest pattern that works.

CrewAI and AutoGen

CrewAI is built around roles and tasks. AutoGen, from Microsoft, is built around agents chatting in conversations.

More agents, more cost

Every agent turn is another LLM call. Three agents can cost several times one agent, so measure if the quality gain is worth it.

See it in code

Walk through the code

python

1from crewai import Agent, Task, Crew2 3researcher = Agent(role="Researcher", goal="Find accurate facts about UPI",4                   backstory="A careful fact-checker who notes sources")5writer = Agent(role="Writer", goal="Explain UPI simply to students",6               backstory="Writes short, friendly blog posts")7 8facts = Task(description="Collect 3 facts on how a UPI payment works",9             expected_output="3 bullet points with sources", agent=researcher)10post = Task(description="Write a 150-word post using the facts",11            expected_output="A 150-word blog post", agent=writer)12 13crew = Crew(agents=[researcher, writer], tasks=[facts, post])14print(crew.kickoff())

1/3Two agents, each with a role, a goal and a backstory that shapes how it behaves.

A two-agent CrewAI crew. Tasks run in order, and the first task's output is passed to the second as context.

Save yourself the pain

Common mistakes

Oops

Starting with five agents because it looks impressive in a demo.

Do this instead

Start with one agent. Split into more only when you can point to a specific failure that a separate role fixes.

Oops

Letting agents chat with no turn limit.

Do this instead

Set a max number of rounds. Two polite agents can thank each other forever.

Your mission

Build a two-agent crew

~40 min

Quick check

Why give the writer agent no search tools?

Go deeper

1Build a simple two-agent system (e.g. a worker and a reviewer) using AutoGen or CrewAI.
2Learn the coordination patterns: sequential, hierarchical, and debate-style.
3Identify when a multi-agent design actually beats a well-prompted single agent — it isn't always worth the added complexity.