Big news! 
We’re launching n8n Agents: a new way to build agents. You define an agent once, and use it anywhere: chat with it directly, drop it into a workflow as a node, connect it to Slack, or run it on a schedule.
To be clear about what’s new here: an agent is a new kind of thing in n8n, not to be confused with the existing AI Agent node you may already use inside workflows. Agents live in their own tab, right next to your workflows:
You give an agent instructions, tools, and a model, and it decides the steps itself. It picks up conversations where they left off, and once you’ve created an agent, you can use it everywhere at the same time.
Agents are available in Preview on n8n Cloud from version 2.32.3.
Please update to version 2.34.x to get the latest improvements!
We’re rolling out in phases over the next couple of weeks.
If you don’t see the Agents tab in your project yet, check again after a couple of days and please be patient.Running self-hosted? Agents run on self-hosted too (Beta), from the same version. See Self-hosted setup in the docs. Not available for self-hosted Enterprise yet.
Why we built this
You could already hand a decision to AI inside of a workflow: the AI Agent node handles the step, while the process around it stays fixed. What you couldn’t do well in n8n was the reverse: put the agent in charge of the whole thing, with access to tools and existings workflow.
Both directions are useful, and which one you want depends on the job.
Sometimes you want the control of a workflow around the agent, with the process firmly in charge, and that’s exactly the right solution for a lot of processes.
Other times you want the agent itself to be more self-sufficient: in charge of the goal, holding the conversation, and deciding what to do next.
You could already build that second kind in n8n, but it meant assembling the pieces by hand: a chat trigger for a way in, memory nodes so it remembered anything, a workflow wrapped around an AI Agent node to hold it together. Agents are built for that case: not to fix the AI Agent node, but to open up use cases it was never meant to cover.
With Agents, getting started is a fraction of the work. Open the Agent Builder, give the agent instructions, a model, and tools, and you have something you can talk to. Memory, channels, schedules, and publishing come built in; within the goal you set, the agent decides what to do next.
Defining the agent in one place also makes it reusable. The same agent can answer people in Slack, run on its schedule, and serve five different workflows as their decision step. Update it and publish, and every place it’s connected gets the new version.
Sometimes you want the procedure in charge, with the agent inside it. Sometimes you want the agent in charge, with procedures inside it. n8n now does both, and the two call each other.
What you can build
- Something people talk to, wherever they work: connect it to Slack, Telegram, or Linear; it holds the conversation and picks it up again next time.
- Something that runs on its own: put it on a schedule and let it handle a recurring job, like a daily triage pass or a weekly summary.
- An agent that runs your workflows: give it your existing workflows as tools, and it decides which one to run and when. Between your workflows and n8n’s integrations, it doesn’t start from an empty toolbox.
- A workflow that calls an agent: message an agent from any workflow with the Message an Agent node, and its reply flows on to the next node.
- An agent with a team: attach other published agents as sub-agents, and it delegates the parts a specialist handles better.
What you need to know to get started
Before you start
- First, check you have access. You need n8n version 2.32.3 or higher, and we’re rolling out gradually: if you don’t see the Agents tab in your project yet, it’s on its way.
- On n8n Cloud, Agents are included in your plan. Not on Enterprise yet.
- You’ll need a model credential. Agents run on your own model account: pick a provider and model when you set the agent up, add the credential when prompted, and token costs go to your provider directly.
- Pricing on the n8n side: one turn with an agent counts as one execution. Tool calls to your workflows and calls to sub-agents don’t count separately, and there’s no AI-credit meter for talking to your agent. Agents draw on the same execution quota as your workflows.
Building your first agent
The best way to build an agent: tell the AI Assistant what you want it to do. It drafts the name, instructions, tools, and skills, and you refine from there.
Prefer to build by hand?
Open the Agents tab, select Create Agent, and set up the pieces yourself:
- Name and model. Pick a provider and model; n8n prompts you for credentials if needed.
- Instructions. The agent’s role, tone, what it should and shouldn’t do, and which tools to prefer. If it misbehaves, fix the instructions before adding more tools.
- Tools. Built-in n8n integrations, your own workflows, custom tools from a JSON schema, or external tools via MCP servers.
- Preview, then publish. Chat with the draft to check its behavior, then publish to make it live. Channels and schedules only run the published version.
Worth knowing as you build
- Skills bundle instructions with the tools needed for one specific task. Reach for them when your agent handles several distinct jobs.
- Knowledge: upload files (csv, pdf, markdown, txt) for the agent to search and read. Available on n8n Cloud; on self-hosted it needs a sandbox (see the docs).
- Memory: the agent keeps conversation context by default, no setup needed. To let it remember across conversations, enable episodic memory in settings (needs an OpenAI credential).
- Channels: Slack, Telegram, and Linear. Each has its own connection and permission setup.
- Schedules: hourly, daily, weekly, monthly, or a custom cron expression.
What it looks like inside of a workflow
Message an Agent node:
Choosing an agent
Inside the canvas
Clicking on the node
You stay in control
- Draft and published versions. Your edits go to the draft; what’s live keeps running until you publish again. Publish history lets you restore or revert, and Unpublish takes the agent offline while keeping the draft.
- Approvals on sensitive tools. Mark a tool as requiring approval and the agent pauses for your Approve or Reject before it runs, then picks up where it left off.
- Full session history. Every conversation is stored as a session: the messages, the tools the agent used, and any pending approvals, all in the Sessions tab.
- Credentials stay in n8n. Each tool runs with the credential you attach to it. Add API keys and passwords in the credential screens, never in the chat.
It’s a Preview - stuff you should know
- Agents can make mistakes and their behavior may change while the feature is in development. Test with Preview before publishing, and use approvals on anything sensitive.
- The existing AI Agent node keeps working, and so do the workflows you’ve built with it.
- Some things people ask about aren’t here yet: agents that watch an event stream continuously, full multi-agent orchestration, and being callable by agents outside n8n. They’re on our minds, but don’t plan around them.
Things to try
- “Create a support agent that answers questions from the PDFs I upload, and escalates to Slack when it can’t answer.”
- “Build an agent that checks Linear every morning for issues with no assignee and posts a summary to Slack.”
- “Give this agent my existing ‘Enrich Lead’ workflow as a tool, and have it use that before replying.”
FAQ
Some questions you might be asking yourself …
Q. How does this relate to the AI Assistant launched a few weeks ago?
A. The AI Assistant is how you build things by describing them. An agent is one of the things it can now build for you, alongside workflows. The quickest way to make an agent is to describe it to the AI Assistant and let it draft the configuration, though the Agent Builder works on its own if you’d rather do it by hand.
Q. How is an Agent different from the AI Agent node I already use?
A. The AI Agent node you know is defined inside one workflow: its instructions and tools live in that node, and it exists only while that workflow runs. A new Agent is defined once, in its own builder, with its own memory and its own home in your project. Workflows can message it with the Message an Agent node, and the same agent can also answer people in Slack, run on a schedule, and serve other workflows, all at the same time.
Q. What does this mean for the AI Agent node I’m already using?
A. Nothing you need to act on. The AI Agent node stays as it is, your workflows keep working, and there’s nothing to migrate or prepare for. Over time we want Agents to become the best way to build this kind of thing, and if that ever means changes to the existing AI Agent node, you’ll get plenty of notice.
Q. Do agents and workflows work together?
A. Yes, in both directions. Your workflows become tools the agent can call, and a workflow can message a published agent with the Message an Agent node. Nothing you’ve built is deprecated.
Q. Can I use my own model and API key?
A. That’s how it always works: agents run on your own model credential. You choose the provider and model, and token costs go to your provider rather than to n8n AI credits. (Building the agent with the AI Assistant is a normal AI Assistant conversation, so that part uses AI credits as usual.)
Q. What does an agent remember?
A. Out of the box, it keeps the context of the conversation it’s in, and you can reopen any past conversation from the Sessions tab and continue it. If you want the agent to carry what it learned across conversations, turn on episodic memory in its settings.
Q. How is this priced?
A. One turn with an agent is one execution. A turn is one exchange: you send a message, the agent responds. Workflow-tool calls and sub-agent calls inside that turn don’t count separately. Agents share your workflow execution quota.
Q. Does it work on self-hosted?
A. Yes, from version 2.32.3. Building agents manually just needs the agents module enabled; AI-assisted building and knowledge bases take a bit more setup. The docs cover both paths: Self-hosted. Self-hosted Enterprise isn’t supported yet.
Q. Can other systems or external agents call my agent?
A. Not yet. Right now agents are reached through the Agent Builder chat, the channels listed above, schedules, or your own workflows.
Q. How is this different from just building the same thing with the AI Agent node and a chat trigger?
A. You could get partway there. What you’d be missing is the part that makes it an agent rather than a workflow: an identity people and workflows can address by name, one definition you can reuse across as many workflows as you like, built-in memory and session history in one place, versioning with draft and publish, and channels and schedules out of the box.
Q. I don’t see the Agents tab. What do I do?
A. Check you’re on version 2.32.3 or higher. If you are, you’re likely waiting on the phased rollout. It’s coming.
We’d love your feedback
This is an early Preview, and your feedback directly affects what we build next. Reply to this post with:
- What you built and how it went
- Where the agent did something you didn’t expect
- What you tried to build and couldn’t
- Which channels, tools, or triggers you want next
Read more in the docs: Build and manage agents
Thanks for building this with us in the open! 
