How Do You Build Your First n8n Agent?
n8n now has an Agents tab, and one agent turn counts as one execution, even when the agent calls several of your workflows.
AI Stella and Liam are AI hosts with AI-generated voices. About our AI hosts
Key takeaways
- n8n Agents live in their own Agents tab, next to your workflows, and the AI Agent node keeps working.
- Create an agent with a name, a model, clear instructions and tools, then Preview it before you Publish.
- Channels and schedules only run the published version, so your draft can stay messy.
- Mark any tool that sends or changes something as requiring approval.
- One agent turn counts as one n8n execution, and model tokens are billed by your own provider.
Show notes
n8n now has an Agents tab, and one agent turn counts as one execution, even when the agent calls several of your workflows. In this episode of Agentic Pod, Liam and Stella show how to build your first n8n agent, run it from Slack or on a schedule, keep it safe with approvals, and watch the two meters that decide your bill.
What we cover:
- What n8n Agents are, and how they differ from the AI Agent node
- Create an agent: model, instructions, tools, Preview, Publish
- Channels (Slack, Telegram, Linear) and schedules
- The Message an Agent node inside your workflows
- Approvals for tools that send or change things
- Billing: one turn is one execution, model tokens go to your provider
- Preview limits and the honest verdict
Chapters
Questions answered in this episode
What are n8n Agents?
A new kind of object in n8n, in its own Agents tab. You give an agent instructions, a model and tools, and it decides its own steps. Your workflows can be its tools.
How do you build your first n8n agent?
Open the Agents tab and select Create Agent, pick a model, write the instructions, add tools, test the draft in Preview, then Publish.
How does n8n bill agents?
One turn with an agent counts as one execution, and workflows or sub-agents it calls in that turn are not counted separately. Model usage is billed by your own model provider.
What is missing in the preview?
Agents that watch an event stream non-stop, full multi-agent orchestration, calling an agent from outside n8n, and queue mode support.
Links & resources
Full transcript
Intro
Hey hey hey, welcome to Agentic Pod! I'm Liam.
And I'm Stella! Liam, n8n added an Agents tab, and one agent turn counts as one execution. Even if it calls several workflows.
Yep. You, uh, build it once, then use it from Slack, on a schedule, or inside a workflow.
Okay, but, um, I already have workflows. Why do I want a second thing?
Good question. Let's start with what it actually is.
What n8n Agents actually are
So is this the AI Agent node with a new coat of paint?
No. The AI Agent node still works, and n8n says it isn't deprecated. Agents are a new object, in their own tab next to your workflows.
And what's different?
You give it instructions, a model and tools. Then it decides its own steps. A workflow follows your map. An agent picks the road.
Ooh. So it's like, a GPS versus a printed map.
Exactly. And your existing workflows can be its tools.
Wait, so it runs my workflows on its own?
It chooses which one, and when. That's the useful part.
Is it ready for everybody?
It's a preview, from version 2.32.3. Keep that word in mind. We'll come back to it.
Build your first agent
Okay, steps. I skip docs, remember?
Open the Agents tab and hit Create Agent. Or describe it to the AI Assistant and let it draft a first version for you.
Then what?
Pick a name and a model. Add your credential for that provider. Then write the instructions: role, tone, dos and don'ts.
And tools?
Built-in integrations, your own workflows, custom tools from a JSON schema, or tools from an MCP server.
Cool. So I add all the tools right away?
No. I mean, if it misbehaves, fix the instructions first. Then add tools.
That's good advice. For life, honestly.
Next is Preview. You test the draft in chat, and only then you hit Publish.
Why does publish matter?
Because channels and schedules only run the published version. Your draft can be a mess and nothing live breaks.
Slack, schedules and approvals
Okay, so how do I actually use it? Just chat inside n8n?
You can. Or connect a channel: Slack, Telegram or Linear. Each one has its own connection setup.
And if I want it running while I sleep?
Schedules. Hourly, daily, weekly, monthly, or a custom cron.
And my old workflows?
There's a Message an Agent node. A workflow sends a message, the agent replies, and the reply goes to the next node.
Hmm. So, like, a robot with a schedule and my email access. What could go wrong?
That's what approvals are for. Mark a tool as needing approval, and the agent pauses until you hit Approve or Reject.
Okay, that's the first switch I'd flip.
Same. Now, billing. It runs on two meters.
Gotchas and honest verdict
Two meters. Go.
One turn with an agent counts as one n8n execution. Workflows or sub-agents it calls in that turn aren't counted separately.
And the second meter?
The model. Agents run on your own credential, so token costs go straight to your provider.
So cheap in n8n, maybe not at the model company.
Right. Yeah, watch both. And remember, it's a preview, so agents can make mistakes and behavior may change.
What's missing?
Agents that watch an event stream non-stop, full multi-agent orchestration, and calling an agent from outside n8n. Queue mode isn't supported yet either.
And memory?
By default it keeps the current conversation. To remember across conversations you turn on episodic memory, which needs an OpenAI credential.
Okay. Verdict. Um, who should try it?
If you already live in n8n, try one low-risk agent this week. Say, a Slack helper that summarizes something. Keep approvals on.
And who should wait?
Anyone who needs queue mode, or an agent that other systems can call.
Fair.
Quick recap. Write the instructions first, preview, then publish. Watch both meters. Links to the guide and the sources are in the description. What's the first job you'd hand an n8n agent? Tell us in the comments. See you next time!