I see abstract AI agent architectures everywhere.

But no one explains how to build them in practice.

Here's a practical guide to doing it with n8n.

1. Single Agents

Selected variants:

• Using tools

• Mixing tools with MCP servers

• With a router (a fancy name for a condition)

• With a human in the loop (Slack approval)

• Dynamically calling other agents

2. Multiple Agents

Selected variants:

• Working sequentially

• Hierarchy with parallel execution and shared tools

• Hierarchy with a loop and shared RAG

3. Best Practices

Here’s what works best for me:

• Ask the agent to plan its work, pursue the plan until the objective is met, and reflect after each iteration.

• Add memory so the agent can track its progress.

• Use a loop to better control complex processes.

• Suggest common tool usage patterns in the prompt (e.g., the order).

• Make sure tools and MCP servers have clear descriptions.

• Check “Return Intermediate Steps” in the Agent settings to debug the thought process.

• Select “Error Workflow” in the workflow settings to handle exceptions.

• If you're using the community version without global variables, create a dedicated workflow to get values by variable name instead of hardcoding them.

• Clearly assign roles and objectives (e.g., planner, researcher, reviewer).

Learn by building, not theorizing.

🎁 You can download my poster as an n8n workflow definition (json, Google Drive): drive.google.com/file/d…

Hope that helps!

P.S. You can download 50+ PM infographics for free by subscribing here: