What are n8n Agents?
n8n Agents are a new way to build AI agents in n8n, launched on September 25, 2026. You describe what the agent should do, pick a model, and give it the tools and workflows it may use; the agent works out the steps. Memory, sessions, channels, versions and approvals come built in, so you no longer assemble them node by node.
What an n8n Agent bundles
The announcement is n8n's blog post Introducing n8n Agents, and the feature list is in the n8n release notes. In n8n's words: "Memory, sessions, channels, versions and approvals come with every agent, so you spend your time on what the agent should do."
Each agent lives in its own Agents tab as a project object, separate from workflows. The existing AI Agent node inside workflows is unchanged, so nothing you already run breaks.
How do n8n Agents differ from the AI Agent node?
The AI Agent node is one step inside a workflow that you wire up yourself, including memory and any approval step. An n8n Agent is a standalone object with those parts already attached, and it can be reached from several places at once: Slack, a schedule, or any workflow through the new Message an Agent node.
| AI Agent node (in a workflow) | n8n Agent (new) | |
|---|---|---|
| Where it lives | Inside one workflow | Its own Agents tab, reusable |
| Memory and sessions | You add and configure them | Built in |
| Approvals | You build a human-in-the-loop step | Built in |
| Versions | Workflow versioning | Published version plus an editable draft |
| Entry points | The workflow's trigger | Slack, schedule, Message an Agent node |
| Tools | Nodes connected to the agent | n8n nodes, existing workflows, MCP servers, sub-agents |
n8n also added "skills": reusable instructions and reference files that several agents can share. For a business with a few agents, that is where the house rules go, such as how to greet a lead or which questions qualify one.
What do n8n Agents cost?
n8n says one turn with an agent counts as one execution, and tool calls to your workflows and to sub-agents do not count separately. The model itself is paid through your own model credentials or n8n's Gateway credits. Agents are available to everyone on n8n Cloud and to self-hosted users with extra setup; Enterprise support is listed as coming soon.
For budgeting, that means the execution count stays predictable, and the variable cost is the model. A qualification agent that runs a short conversation per lead costs a few model calls per lead. Check the model's own pricing before you pick one; a smaller model is often enough for routing and tagging.
Are n8n Agents ready for production?
n8n describes Agents as in preview: "Everyone on Cloud can use agents and they work. We're improving them release by release, so test before publishing." For customer-facing jobs we would run a new agent beside the existing workflow first and compare results before switching.
That matches the wider evidence. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, mostly on cost, unclear value and weak risk controls (see our AI agent adoption research). The fix is scope, not a new tool: one agent, one job, with an approval step for anything that spends money or messages a customer.
When should you use an agent instead of a workflow?
Use a plain workflow when the steps are known in advance: form submitted, contact created, text sent, rep notified. Use an agent when the next step depends on reading something unstructured, such as an email reply, a support ticket or a free-text enquiry, and deciding what to do with it.
Workflow or agent?
- Workflow: syncing a new lead from a form to your CRM, sending reminders, creating invoices. Cheaper, faster, and easy to audit.
- Agent: reading inbound emails and deciding whether each is a new lead, a support request or spam; drafting a reply for approval; chasing missing documents with judgment about what is missing.
- Both: a workflow handles the trigger and the CRM writes, and calls an agent through the Message an Agent node only for the step that needs judgment.
The third pattern is the one we recommend for most n8n automation builds, because it keeps the predictable parts predictable. If you are weighing n8n against other tools, our Zapier automation and API integration pages cover when each fits.
What should current n8n users do?
Nothing is forced. Existing workflows and AI Agent nodes keep working. If you have an agent built from several nodes with hand-made memory and approvals, it is worth rebuilding one as an n8n Agent in a test project and comparing behaviour and execution counts before moving anything live.
Moving one AI step to an Agent
- Update to the latest stable n8n version (Cloud users already have it).
- Pick one existing AI step with a clear success measure.
- Rebuild it as an Agent with the same model and tools.
- Run both on the same inputs for a week and compare.
- Switch only if the agent is at least as accurate, and keep the approval step.
Frequently asked questions
Are n8n Agents available on self-hosted n8n?
Yes, with additional setup, according to n8n's announcement. Cloud users get them on the latest stable version.
Do n8n Agents replace the AI Agent node?
No. The AI Agent node still works inside workflows. Agents are a separate, higher-level option.
Can an n8n Agent talk to customers?
Agents connect to channels such as Slack, Telegram and Linear, and can be called from any workflow, so a workflow can pass a customer message to an agent. Keep an approval step for anything sent to a customer while the feature is in preview.
See our AI workflow automation service, the case studies, or contact us to talk through a specific agent. More updates are on the news page.
