n8n released Agents on 25 September, a goal-driven way to build on a workflow automation platform that has always asked users to lay out every step first. Two design choices underneath the launch deserve more attention than the plain-language builder does: where the credentials live, and how a conversation gets billed.
Key takeaways
- An n8n agent is assembled from instructions, a model and a tool list, then reached from Slack, Telegram, Linear, Discord, a schedule, or a node inside an ordinary workflow — one published agent behind all of them.
- Credentials attach to each tool rather than to the agent, so a model that can append a CRM note never holds the CRM key.
- A turn with an agent costs one execution from the same quota workflows already draw on, with calls into workflow tools and sub-agents not charged on top.
How the permission boundary works
Tools come in three shapes. An MCP server hands over its whole catalogue at once and the agent works out which call it needs, with unwanted tools excluded at setup. A single n8n node gives one action with parameters fixed in advance. A workflow gives a process that runs exactly as it was built.
That third option carries the security argument. n8n's own example is a support agent whose note-taking tool is a workflow that appends text to an account and can do nothing else; the alternative is granting a model write access to the CRM and trusting its instructions to stay inside the notes field. Tools flagged as sensitive stop for a human approval before they fire, per-tool credentials keep instance keys away from the model, and editing or publishing rights follow whatever n8n roles an organisation already has. n8n's own advice for anything sensitive is to start where the blast radius is small — scoped tools, a test channel, approvals on writes.
What changes for teams already on n8n
Nothing breaks. The AI Agent node stays exactly as it was, so the assembled pattern many teams run today — chat trigger, memory node, tools attached, a workflow holding it together — is unaffected. Agents are a separate object placed beside workflows, and traffic flows in both directions: a new Message an Agent node lets a fixed workflow hand one step to the agent and carry the answer onward.
Operationally, the useful detail is versioning. A draft can be edited and previewed while the team keeps using the published agent, and publishing updates every channel and every workflow that references it at once. Sessions store each conversation with its tool calls and their inputs and outputs, which is the difference between an agent you can audit after a bad Tuesday and one you cannot. Memory persists within a session by default and optionally across them, skills hold reusable instructions that several agents can share, and uploaded CSV, PDF, Markdown or text files ground answers on n8n Cloud.
What a turn actually costs
Metering is where n8n has drawn a line that cuts against the market. Because one turn is one execution and internal tool calls are not counted separately, the cost of an agent conversation does not scale with how many steps the model decides to take — a meaningful difference from pricing that charges per action. Gateway credits let a team run a first agent without bringing a provider API key, while building one through n8n Assistant draws on AI credits like any other Assistant session.
Where this lands in a crowded agent market
Every automation vendor is shipping an agent builder this year, and most of them start the buyer at zero with a catalogue of generic integrations. n8n is starting somewhere else. Its pitch is that the workflows a company has already built and already trusts become the agent's hands, with the scoping work done years ago still in force. Whether that advantage holds depends on execution: the feature ships as a preview, it is on for everyone running the latest stable n8n Cloud release, and self-hosted instances need extra configuration before agents will run at all. The company's own framing — following an assistant earlier this year that builds a workflow and then debugs it — suggests the canvas is not being replaced so much as demoted to one option among two.
FAQ
Does the existing AI Agent node still work?
Yes. n8n says the node is unchanged and everything built with it keeps running. Agents are a new object type that sits alongside workflows rather than a replacement for the node, and an AI agent and a workflow can invoke each other.
Can self-hosted n8n users run Agents?
Yes, with extra setup. The feature reached everyone on n8n Cloud running the latest stable release at launch, while self-hosted deployments require additional configuration covered in the n8n documentation. Knowledge file uploads are listed as a Cloud capability.
How is a conversation with an n8n agent billed?
One turn equals one execution against the standard workflow allowance. The agent's calls into workflow tools or sub-agents are not billed separately, though creating an agent through n8n Assistant consumes AI credits.
