3.5 — Demo: ReAct Loop 🧪

Step through the Reason → Act → Observe → Loop cycle (until it can answer) on an example mission, step by step, then visually compare the 3 multi-agent paradigms (graph / roles / conversation — detailed right after, in 3.6). What you observe here is exactly what Make or n8n run for you in 3.8.

ReAct Loop— Reason → Act → Observe

The missionWhat's the weather in Paris today, and do I need an umbrella?

🧠

Reason

I need the weather in Paris. I'll use my web search tool.

Next step →1 / 5

One agent, or a team? The 3 paradigms

When a single agent is no longer enough, three ways to make agents collaborate:

LangGraphgraph · strict control

Plan→Tool→Check→End

Graph flow, each step controlled. State memory (checkpointing): resumes after a crash. The production default.

CrewAIroles · intuitive

Manager↓Researcher↓Writer

A team with roles, like in a company. The manager delegates, each has a specialty.

AutoGen / MAFconversation · debate

Coder : Here's my function.

Tester : Fails on one case — try again.

Coder : Fixed ✓

Agents debate in a thread. Effective for code. AutoGen merged into Microsoft Agent Framework (late 2025).

⚠️ Educational simulation in your browser (no real AI called). ReAct is the mental model; frameworks automate this loop for you.