n8n Agents are here: when to use an agent vs a workflow
n8n now has a first-class Agent product alongside workflows. The important change is not another AI node—it is a new way to decide what stays deterministic and what the model can choose.
n8n has added a first-class Agent product alongside its workflow builder. The practical change is bigger than adding another AI node: teams can now decide which parts of an automation should stay deterministic and which parts can be left to a model to work out at runtime.
What changed
An n8n Agent is configured around a goal rather than a fixed canvas. You choose a model, write instructions, attach channels and triggers, and give it tools. Those tools can be MCP servers, individual n8n integrations, or complete workflows you already trust.
Workflow or Agent?
| Use a workflow when… | Use an Agent when… |
|---|---|
| The sequence is known in advance. | The next step depends on what happens during the task. |
| A write action needs tightly controlled inputs. | The job is conversational or open-ended. |
| Consistency and predictable execution matter most. | The model needs to choose among tools or ask follow-up questions. |
| You want a fixed, auditable process. | You want one reusable agent across Slack, schedules and workflows. |
The useful part: agents can call workflows
The strongest design pattern in the release is not giving an agent unrestricted access to every system. It is letting the agent decide when a task is needed, then handing the sensitive action to a narrow workflow that performs one controlled operation. n8n’s own example uses a workflow that can add a CRM note without exposing broader CRM write access to the agent.
This creates a clean control boundary: let the agent own ambiguity, but let workflows own irreversible or highly structured actions. For a sales agent, for example, researching a lead and deciding what context matters can be agentic; updating deal fields, sending an approved email, or changing a record can remain a constrained workflow with explicit inputs.
What comes built into an n8n Agent
- A model plus plain-language instructions.
- Channels and triggers including Slack, Telegram, Linear, Discord and schedules.
- Tools from MCP servers, n8n integrations and existing workflows.
- Reusable skills and sub-agents.
- Knowledge files and memory across sessions.
- Stored sessions with tool-call and execution visibility.
- Draft and published versions so changes can be tested before replacing the live agent.
Existing n8n AI Agent workflows do not need a migration
n8n says the existing AI Agent node remains unchanged. If your current agent is already well-defined inside a workflow, there is no automatic reason to rebuild it. The new Agent product is more compelling when the agent itself should be reusable across multiple channels or workflows, needs persistent sessions, or should choose among a larger set of tools.
Cost and availability
- n8n says one turn with an Agent counts as one execution.
- Calls from that Agent to workflows and sub-agents do not count as separate executions.
- Agents are available on n8n Cloud for users on the latest stable version.
- Self-hosted Agents are available with additional setup.
- Enterprise availability is listed as coming soon.
- Agents are still in preview, so n8n recommends testing before publishing and using approvals for sensitive actions.
A safer rollout pattern
- Start with a low-risk internal use case such as research, triage or a morning summary.
- Give the Agent only the tools it needs instead of broad system access.
- Wrap sensitive writes in narrow workflows with explicit parameters.
- Require human approval for actions with meaningful business impact.
- Review stored sessions and execution logs before expanding the Agent’s permissions.
- Only then move the same published Agent into more channels or production workflows.
Agents do not make deterministic workflows obsolete. n8n explicitly supports both directions: an Agent can call a workflow, and a workflow can call an Agent. The useful architecture is often a hybrid—discretion where the task is ambiguous, deterministic automation where the action must be controlled.
n8n Agents turn the platform into more than a workflow canvas: the agent can now be the reusable unit that carries instructions, memory, tools and sessions across channels. The key design decision is not agent versus automation. It is deciding exactly where the model gets discretion and where a workflow should keep control.
Sources & useful resources
- n8n — Introducing n8n Agents— Primary announcement and product details.