n8n gives you two ways to build AI agents: the AI Agent node, which sits inside a normal workflow and needs a chat model plus at least one tool, and the newer Agents feature, which builds a standalone assistant people reach through chat, channels and schedules. As of October 2026 the docs mark Agents as Preview, so the AI Agent node is the steadier place to start.
If you've never built an n8n workflow, read how n8n works for beginners first, because an agent is just one part of a workflow. This page is only about the agent part: what it needs, when it's the right tool, and how to stop it doing something you'd regret.
What makes something an agent in n8n
n8n's docs put it neatly: an agent is a chain that knows how to make decisions. A chain follows a predetermined sequence of calls to AI components. An agent uses a language model to decide which actions to take, and which tools to use to answer the request in front of it.
Here's the detail most people don't expect. When a workflow with an agent runs, the agent runs several times. The docs give an example: an initial setup, then a run to call a tool, then another run to evaluate the tool's response and reply to the user. That's why one agent step can feel slower, and do more model calls, than the single box on your canvas suggests.
Two routes: the AI Agent node and Agents
| Route | Where it lives | Status as of October 2026 | Best for |
|---|---|---|---|
| AI Agent node | Inside a workflow, with a chat model and tools attached as sub-nodes | Established. The old agent type setting is deprecated from n8n 1.82.0 | One decision-making step inside a bigger automation you already run |
| Agents | Its own Agents tab in your project, next to your workflows, built in n8n's own editor (also called Agent Builder) | Preview. On n8n Cloud on all plans, and self-hosted on all plans from n8n 2.32.3 | Open-ended requests from people, reached through chat, Slack, Telegram, Linear or Microsoft Teams, or run on a schedule |
Don't let the name confuse you: n8n's Agent Builder is n8n's own editor for its Agents feature, not OpenAI's product of the same name. The docs also say plainly that Agents can make mistakes and their behaviour may change while the feature is in development. Check the current docs before you build anything important on it.
What each route needs before it works
For the AI Agent node, the rule is short. Connect a chat model and one or more tools, and the agent decides which tools to call. You must connect at least one tool sub-node; it isn't optional. n8n connects to model providers such as OpenAI, Anthropic and Google, lets you add tools and memory, and can connect an agent to an MCP registry server in one click.
The Agents feature has more parts. Each agent has a model, instructions (its system prompt), and capabilities you add: tools, web search, skills, a knowledge base, memory and sub-agents. Channels and schedules only take effect once you publish. Three details worth knowing before you start:
- Publishing takes a snapshot of your draft. Edits after that stay in the draft until you publish again, so the live agent doesn't change while you tinker.
- Session memory is on by default, but episodic memory, which recalls earlier conversations, needs an OpenAI credential to store and retrieve memories.
- Knowledge files can be csv, pdf, markdown or txt. On self-hosted n8n, knowledge bases are also in Preview.
The docs' advice on instructions is the one I'd pin to the wall: if the agent doesn't behave as expected, refine the instructions before you add more tools.
Agent or chain? Pick the dull option when you can
Not every AI step needs an agent. The Basic LLM Chain node lets you chat with a connected model, but it doesn't support memory or tools. That sounds like a weakness. For a lot of business jobs it's a strength: summarise this email, classify this lead, rewrite this product description. The steps are known, so a fixed workflow does them the same way every time.
My rule of thumb: if you can write the steps down, use a fixed workflow or a chain. Use an agent only when the task is open-ended enough that you can't predict which tool it needs. And whichever you pick, add an approval step on anything that sends or deletes.
Put a human in the loop before it sends or deletes
n8n lets you require approval before an AI Agent runs a specific tool. When the agent wants that tool, the workflow pauses and a person gets a request showing the tool and the details the AI wants to use. They can Approve, and the tool runs with the AI's input, or Deny, and the action is cancelled. The docs suggest it for tools that do irreversible things, such as deleting data, sending external messages or making purchases. Setting it up on the AI Agent node:
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
- In your workflow, click the Tools connector on the AI Agent node to open the Tools Panel.
- Find the Human review section and pick your approval channel.
- Configure that channel with its credentials and settings.
- Add the tools that need approval to the human review step's tool connector, then set each tool up as normal.
- Tell the agent in its system prompt which tools need approval and what to do when a request is denied.
The approval channel can be n8n's Chat, Slack, Discord, Telegram, Microsoft Teams, Gmail, WhatsApp Business Cloud, Google Chat or Microsoft Outlook, and it can differ from the channel your users talk to. In the Agents feature the same idea shows up as Approve or Reject buttons in the chat. n8n's own advice is to start with human review on and reduce oversight as your confidence grows. I agree.
The version catch in older templates
If you're copying an agent workflow from an old tutorial, check the node version. The AI Agent node's agent type setting is deprecated from n8n 1.82.0, and every AI Agent node now works as a Tools Agent. Version 1 of the node, the one with that setting, will be removed in n8n 3.0, so update it if that's what you're running. The older types, Conversational, OpenAI Functions, Plan and Execute, ReAct and SQL, are all listed as deprecated. For an SQL-style job, the docs suggest a Postgres or MySQL tool sub-node on a recent Agent node instead.
Where I'd start
I'd start with one AI Agent node inside a workflow you already trust, with a single tool and human review switched on. Skip the Agents feature for anything customer-facing until it leaves Preview, unless you're happy to keep checking the docs for changes. If you're still choosing between automation tools, my n8n and Zapier comparison covers that decision first.
If you'd like help deciding where an AI agent belongs in your workflows and where a fixed automation is safer, that's the kind of workflow automation I help with. I run AI training and workshops for teams, and help businesses with AI enablement and workflow automation, so the busywork runs itself and people get time back. Most of that work is with SaaS marketing teams and small businesses. Here's how we can work together.
Frequently asked questions
Does n8n have AI agents?
Yes. n8n has the AI Agent node, which works inside a workflow, and a newer Agents feature for standalone agents, which the docs mark as Preview as of October 2026.
What does the n8n AI Agent node need to work?
A chat model and at least one tool. The docs say you must connect at least one tool sub-node to an AI Agent node, and the agent then decides which tools to call.
What's the difference between an agent and a chain in n8n?
A chain follows a predetermined sequence of calls. An agent uses a language model to decide which actions to take. The Basic LLM Chain node doesn't support memory or tools.
Can an n8n agent ask for approval before it acts?
Yes. You can require human review before the AI Agent runs specific tools. The workflow pauses and a person approves or denies the request through a channel such as Slack, Gmail or n8n's Chat.
Which AI models can n8n agents use?
n8n connects to model providers such as OpenAI, Anthropic and Google. You pick the provider and model when you set up the agent and add credentials if it asks.
