Phase 5 · Multi-Agent Workflows Compared: Why They're "The Same Graph"

Goal: a head-on comparison of mainstream frameworks like CrewAI / AutoGen, and showing they are isomorphic to LangGraph at the level of "graphs". This is the first official appearance of this project's central claim.

Two framework designs, up close

CrewAI: roles + pipeline (crew)

• You define several Agents (roles), each with role / goal / backstory;
• You chain them into a sequential/parallel pipeline (crew) using Tasks;
• The output of one Agent automatically becomes the input of the next.

Its mental model: a few specialists lined up in a pipeline.

AutoGen: conversation + coordinator

• You define several ConversableAgents;
• A GroupChat / coordinator decides "who speaks next";
• Roles message each other until someone says "task done".

Its mental model: a group chat moderated by a coordinator.

Translating them back into graphs

Now re-read them through the lens of previous chapters:

Framework concept	Graph equivalent
Agent / role	node
Task chaining / messaging	edge
CrewAI sequential execution	fixed sequential edge
AutoGen coordinator "who speaks next"	conditional edge
shared state / chat history	graph state

Conclusion: CrewAI's sequential crew ≈ a "linear graph"; AutoGen's group chat ≈ a "graph with conditional edges". Both are fundamentally "state + nodes + edges", just with different APIs.

This is the project's central claim: the underlying logic of mainstream Agent frameworks maps onto LangGraph's "state graph + nodes + conditional edges" model. Master LangGraph, and you master their common language.

Example 1: hand-rolled crew (CrewAI analog)

File: examples/p5/mini_crew.py

Without installing CrewAI, we chain "Researcher → Writer → Reviewer" with three prompt templates. You'll see it's the same structure as a CrewAI crew — CrewAI just automates "feed the previous output to the next step".

python -m examples.p5.mini_crew

Example 2: coordinator routing (AutoGen conditional-edge analog)

File: examples/p5/agent_router.py

A coordinator reads the user request, outputs "route to math / language / code", then executes the chosen node. That routing moment is the conditional edge of AutoGen's coordinator deciding "who speaks next".

python -m examples.p5.agent_router

Key insight

The unification thesis is now almost self-evident:

• Phase 3: single-Agent loop = graph (think/tool/observe nodes + conditional edge)
• Phase 4: tool-using Agent = graph (think node + tool nodes + conditional edge)
• Phase 5: multi-Agent frameworks = graph (role nodes + sequential/conditional edges + shared state)

Since they're all graphs at heart, why not express them uniformly with a framework built for graphs? That's exactly what Phase 6 answers — LangGraph takes the stage.

Acceptance checklist

• [ ] Can describe the design difference between CrewAI (roles + pipeline) and AutoGen (chat + coordinator)
• [ ] Can map "role / message / coordinator decision" to "node / edge / conditional edge"
• [ ] Have run Example 1 (hand-rolled crew) and Example 2 (coordinator routing)
• [ ] Can restate the central claim: "mainstream Agent frameworks are graphs underneath"

Next

Phase 6: LangGraph graph orchestration — using state graphs, nodes, and conditional edges to express everything from earlier chapters in one unified way, and comparing the code-size difference between "framework graph" vs "hand-rolled graph". The unification thesis closes its loop here.

Extra Example

Two-agent debate (examples/p5/debate.py)

One shape of an AutoGen group-chat: two roles take turns challenging each other, driven by a simple turn loop.

python -m examples.p5.debate