In an official blog post, the n8n platform announced the launch of Agents. According to the announcement, users can describe what an agent should do, provide it with a model, tools, and workflows, and the agent works out the steps required to complete the task itself. Users can communicate with the agent in Slack, run it on a set schedule, or call it from any workflow, with it being the exact same agent across all these access points. n8n notes that agents are designed to make it easier to handle open-ended tasks or multi-step back-and-forth conversational jobs that could get complicated to build as a fixed workflow, and users can build an agent without prior knowledge of how workflows operate.
Agents run alongside existing workflows in the system, and the two components were built to work together. An agent can use workflows as tools, allowing the user to determine exactly what the agent is permitted to do across various systems. When a workflow needs an agent for one of its steps, a new node called Message an Agent allows calling it directly from within the workflow. n8n emphasizes that for users already utilizing the existing AI Agent node, nothing has changed and everything built with it continues to work.
Why agents were launched now
According to n8n, since the platform was established, automating something was based on working out exact steps and adding them to the canvas. The company notes that this remains the right approach for many defined tasks, such as a process where a lead enters the system, is enriched with data, scored, and routed to its destination. The more fixed the sequence of actions, the better a structured workflow fits, even if one of the steps includes a model making a decision.
However, two key shifts led to the launch of agents:
- Models are now capable of working out how to act: When a current model is given a goal and the right tools, it is capable of calculating the steps required to achieve that goal.
- People expect to ask for an end result: Users increasingly want to define a goal and receive a solution rather than designing all process steps in advance.
This is especially significant for tasks where the input is different every time. For example, when a team member asks in Slack why a customer's usage dropped last month, answering requires several rounds: pulling account data, clarifying which product line is meant, checking support history, and returning with a summary and a follow-up question. Because the next step depends on the answer to the previous step, the process cannot be laid out precisely in advance. Similar examples include support emails, new GitHub issue reports, or requests where the domain is known but the specific task is not known in advance. Until now, handling such cases in n8n meant fitting an open-ended conversation into a structured workflow, which took work to build.
What goes into each agent in the system
n8n compares the new agents to buying a pre-built computer versus building a PC from individual parts. Previously, one could connect a chat trigger, a memory node, and an AI Agent node with attached tools, but the new agents arrive with built-in components: memory, session management, channels, versioning, and approvals.
An agent setup includes several components:
- Model: Any model for which the user has credentials, or through n8n Gateway credits.
- Instructions: The agent's role, tone, what it is and is not allowed to do, and which tools to prefer using.
- Channels & triggers: Connection to platforms such as Slack, Telegram, Linear, Discord, or running on a schedule.
- Tools: Built-in n8n integrations, MCP servers, and any workflow in the system.
- Skills: Reusable instructions and reference files loaded by the agent when needed, which can be shared across different agents.
- Sub-agents: The ability of one agent to call another agent.
- Knowledge: Uploading csv, pdf, markdown, or txt files to ground the agent's answers (available on n8n Cloud), as well as using vector stores.
- Memory: Retaining the session conversation by default, with the option to remember information across sessions.
- Sessions: Every conversation is stored, reviewable, and can be resumed later. Session execution logs display every step the agent took, which tools it called, and the inputs and outputs of each call.
Using workflows as tools and control mechanisms
An agent in n8n can use three types of tools based on user selection:
- MCP servers: Connecting a service that gives the agent access to all of the service's tools at once, with the ability to exclude specific tools.
- n8n tools & nodes: Using integrations configured for a specific action with chosen parameters, enabling focused control.
- Workflows: An entire process built in advance that runs exactly as defined, without requiring any changes to the original workflow.
For example, a support agent handling incoming tickets can use three workflows as tools: pulling account data from the CRM and checking contract status, adding a note to the account, and paging on-call staff in a dedicated channel. The agent decides when to trigger each workflow, but what occurs within the workflow remains fixed according to the pre-built definition. Through this approach, the agent does not receive direct write access to the CRM, but only access to a workflow that adds a note and nothing else.
Several control mechanisms surround agent operations:
- Approvals: Marking a tool as sensitive causes the agent to pause for human approval or rejection (Approve/Reject) before using it.
- Per-tool credentials: Each tool runs with the credentials configured for it, and the agent never holds keys to the entire instance.
- Edit and publish permissions: Access to the agent is managed according to user roles configured in n8n.
Integrating agents inside workflows and version management
Agents also operate in reverse: when a workflow needs an agent in one of its steps, the Message an Agent node can be added. The node sends the agent a message based on workflow data and passes the agent's response to the next node. The agent brings along its configured instructions, tools, and memory. When agent settings are updated and published, every workflow calling it receives the update automatically.
For team collaboration, the system provides:
- Using the same agent across multiple places (connected to Slack, scheduled, or called from a workflow), where updating instructions and publishing applies to all connections.
- Drafts and published versions: Editing and previewing a draft while the team continues using the published version, with options to restore, revert to a previous version, or unpublish.
- Detailed session logs covering inputs, tool calls, outputs, and errors.
Availability, pricing, and getting started
An agent can be created via the Agents tab by clicking Create Agent, or by describing the task to the n8n Assistant, which determines whether to build a workflow or an agent and drafts the instructions, tools, and channels. Using Gateway credits, users can begin working and select a model without needing an API key from an external AI provider.
In terms of availability and costs:
- Availability: Agents are available on n8n Cloud for all users on the latest stable version, and on self-hosted instances with additional configuration. Support for Enterprise versions is expected to be added soon.
- Cost: One conversational turn with an agent counts as one execution. Calls to workflow tools or sub-agents are not counted as separate executions, and agents share the general workflow execution quota. Building an agent with n8n Assistant consumes AI credits.
- Status: The feature is currently in Preview status, and the company recommends testing before publishing and configuring approvals on sensitive actions.