n8n AI agents

n8n asks a simple question on its AI page (checked Oct 27, 2026): do you need a workflow, an agent, or both?

Here's the difference, in plain English. A regular workflow follows the exact same steps every time (n8n calls this "deterministic"). It doesn't need memory, it runs on a trigger like a schedule or a web event, and it can skip AI costs entirely. An agent adds judgment. It decides what to do as it goes. And in n8n, workflows and agents can call each other in the same workspace.

Here's what n8n says you get: First, you can test how an agent behaves before you publish it, open any step to see what went in and what came out, and replay a run to find where it broke. Next, you can publish in one click, with scheduling, error handling, versioning, and run logs built in. Then, you can control which tools an agent can reach and which actions need a human to OK them. On top of that: you can self-host (run it on your own servers) to keep data in-house, use 600+ native integrations as tools, and share one workspace across teams with access control. You can draft agents with the AI Assistant, through the MCP server (MCP is a standard way to plug AI into other tools), or build them by hand.

n8n's AI Agent node docs (also checked Oct 27, 2026) say you connect a chat model and one or more tools, and the agent decides which tools to call. You need at least one tool connected. The older "agent type" setting is being phased out, and they all work as a Tools Agent now.

n8n's FAQ also explains the difference between an AI Agent node (one step inside a workflow) and a standalone agent that has its own identity and memory between runs.

Pricing: we didn't find clear dollar amounts on the pages we read, so that's unknown for now.