n8n Agents

Models and tools are configurable.

Source: blog.n8n.io
September 25, 2026
8 min

n8n Agents makes agents independent objects alongside workflows: a team configures an agent once, then uses that same agent through chat channels, scheduled tasks, and other workflows. It handles requests with its own instructions, tools, and memory, while the model decides what to do next based on the task and the results so far.

This change addresses how to build for open-ended tasks. Fixed data synchronization still lends itself to a workflow; tasks that require reading material, asking follow-up questions, and then deciding which system to query can be assigned to an Agent. The two approaches can work together, without requiring every process to be handed to a model or every possible conversation branch to be mapped out in advance.

From assembling components to independently managed agents

It was already possible to build agents in n8n by connecting a chat trigger, an AI Agent node, memory, and tools inside a workflow. The new Agents feature brings conversations, memory, channels, versions, and approvals into a single object, reducing that assembly work. The existing AI Agent node remains unchanged, so existing workflows do not need to migrate as a result.

In Agent Builder, the configuration elements serve different purposes:

  • Models and instructions: Choose an available model and define the role, tone, scope of work, and preferred tools. You can use your own provider credentials for the model or choose available Gateway credits.
  • Tools: Provide the ability to take action, such as querying business systems, writing records, or running existing workflows. Text instructions alone, without the corresponding tools, cannot perform these system operations.
  • Skills: Store reusable task instructions and reference material that can be loaded as needed and used across multiple agents.
  • Knowledge: Supply material that can be retrieved when answering questions, with support for CSV, PDF, Markdown, and TXT files, as well as use with vector stores. The launch announcement explicitly identifies n8n Cloud as the environment where file uploads are available.
  • Memory and conversations: Preserve the current conversation context by default and also support cross-conversation memory. Conversations are saved so they can be reviewed and resumed; this serves a different purpose from uploading a set of long-term reference files.
  • Sub-agents: One Agent can call another Agent, delegating a task to another configured agent.

Together, these settings shape how a request is executed. After receiving a message, the Agent can call tools, look up material, delegate tasks, or ask follow-up questions, then continue working with the information it obtains until it produces a reply. It does more than pass input to a model to generate a piece of text.

How tools are connected determines the scope of what it can do

An Agent has three main ways to connect to business systems. The difference lies in how many operational details you define in advance.

MCP servers are useful for connecting a set of capabilities at once. Once a service is connected, the Agent can choose actions from the exposed tools; tools that are not needed can be excluded. n8n's MCP configuration documentation also allows selecting only a subset of tools rather than handing over all of a service's capabilities by default.

n8n integrations and node tools suit clearly defined individual operations. You select the operation and fix the necessary parameters, leaving the model to fill in only the parts reserved for it. When sending a message, for example, you can specify the destination channel in advance instead of letting the model choose it each time. Compared with connecting a set of MCP tools, this requires more configuration but also narrows what the model can change.

Workflow tools let the Agent invoke an entire predefined process. There are two layers here: the Agent decides when to call it, and the workflow determines the steps to execute after the call. Existing business logic can stay in workflows rather than being rewritten into prompts.

The official customer support example illustrates this distinction well: one Agent uses three workflows to retrieve customer context, add customer notes, and notify the on-call team. It chooses tools based on the ticket, drafts a reply, and handles escalation. The context-retrieval workflow queries the CRM, contract status, and customer health, while the note-adding workflow accepts only a customer ID and the note content.

The permission scope for adding notes is particularly important. By using this workflow as a tool, the Agent gains a limited action rather than unrestricted ability to modify the CRM. Each tool uses its own bound credentials; sensitive tools can require approval, and the Agent must wait for a person to select Approve or Reject before continuing. Who can edit and publish an Agent remains governed by n8n's role-based permissions.

The same Agent can respond to people and be called by processes

Chat channels let teams make requests directly, while scheduled tasks let an Agent work on a timetable. The entry points listed in the launch announcement include Slack, Telegram, Linear, Discord, and scheduled tasks. For example, support staff can ask questions in chat while a separate scheduled task has the same Agent organize outstanding items every day, without maintaining two sets of instructions.

Conversely, a workflow can call an Agent through the Message an Agent node: it assembles data from earlier nodes into a message, sends it to a published Agent, and passes the answer to subsequent nodes. The Agent uses its existing instructions, tools, and memory, while the caller primarily handles the connection between inputs and outputs.

A team can therefore place fixed steps in an outer workflow and call an Agent only at the step that requires judgment. Alternatively, it can have the Agent manage the overall task and call several fixed workflows as needed. The choice depends on which parts need a predetermined sequence and which must change approach in response to new information.

Editing, publishing, and investigation belong to one management process

An Agent's draft is separate from its published version. While the draft is edited and tested in preview, the team continues to use the published version; a new version is published when it is ready. Publishing updates the same Agent across its connected entry points, without copying configurations channel by channel. Versions also support restoration, rollback, and unpublishing.

Conversation records contain more than the final answer: they also include execution steps, tool calls with their inputs and outputs, and errors. If something goes wrong, you can inspect which actions the Agent actually chose and what results it received. Logs provide evidence for investigation, while approvals provide a human checkpoint before an action. Neither replaces the other.

Getting started and understanding the different cost layers

To build manually, open the project's Agents tab, select Create Agent, and configure the model, instructions, and capabilities. You can also describe your requirements to n8n Assistant, which can draft instructions, tools, and Skills for you to review and test. Assistant is an entry point for building; the completed Agent is the object that handles business requests.

If you do not have a model provider API key, you can start experimenting with available Gateway credits; you can also use your own credentials. Gateway credits pay for supported model or service calls. They are separate from the AI credits consumed by conversations with Assistant, so the two balances should not be treated as a single allowance.

Agent executions are counted separately: a user sends a message, the Agent processes it and returns a reply, and that round counts as one execution. The announcement states that the Agent's calls to workflow tools and sub-agents do not add separate executions, and that Agents and workflows share the plan's execution allowance. If a conversation involves multiple exchanges, estimating execution volume as just one per conversation would be incorrect; model-call costs are also distinct from this execution count.

This release makes Agents available to n8n Cloud users running the latest stable version and also offers a Community self-hosted option that requires additional setup. Enterprise is in preview, with more governance features still in development. Agents itself also remains a preview feature; the official guidance calls for testing before publishing and retaining approvals for sensitive actions.

For teams with existing workflows, the most valuable starting point is to identify a task that requires repeated queries and judgment, then make existing reliable processes available as its tools rather than rebuilding all automation. Customer support ticket handling, lead research, Linear issue follow-up, and scheduled summaries are all directions suggested in the announcement. Which workflows can be reused, what each tool is allowed to do, and which actions must wait for approval are the key decisions in configuring this Agent as a concrete work assistant.

2026-09-25: Original · n8n

Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.

Original source:blog.n8n.io