n8n turns workflows into tools for AI agents

n8n’s new Agents feature lets a model choose among workflows and connected tools, while keeping credentials and approvals inside the automation platform. The design puts a familiar workflow layer between an AI system and business operations.

On September 25, automation platform n8n introduced Agents, a preview feature that can use workflows as tools. A user describes an objective; an agent can select from connected capabilities, pass information between them and continue until it reaches a result. Administrators can also add MCP integrations and configure approval steps for actions that need a person’s sign-off, according to n8n’s announcement.

The important shift is less the arrival of another chat panel than the placement of agent behavior inside a workflow system. n8n already represents work as connected steps. Its Agents feature makes those flows available as capabilities an AI can choose, while the platform retains the connections and credentials needed to run them.

That can make an agent easier to constrain than a free-form system with broad access to company accounts. A team can expose a workflow that performs a narrow operation instead of handing a model every underlying credential. Approval gates can interrupt selected actions. These are useful control points, but they are not a guarantee: the operator still needs to decide which workflow is safe to expose, what data it can handle, and what happens when the model chooses the wrong tool or supplies bad inputs.

The launch also makes cost and product status visible parts of the design. n8n says the feature is in preview and available to its cloud and self-hosted users. The company counts an agent turn as a workflow execution; individual tool calls within the turn do not each count as separate executions. Building an agent through the assistant can also consume AI credits. Teams evaluating the feature should model those costs against repeated or failed runs, not just the number of successful conversations.

There is a broader platform contest here. Automation vendors already sit near the systems that businesses want AI to operate: ticket queues, databases, messaging tools and internal services. By wrapping established workflows as agent tools, they can sell a path to action without requiring every company to build its own orchestration layer. Inference from the launch: workflow platforms may become the practical permission surface for many workplace agents.

The measure of success will be operational, not conversational. Can administrators inspect what an agent selected, replay or stop its work, limit data exposure, and explain the result after an incident? n8n’s announcement describes building blocks, not independent reliability evidence. For now, it is a preview worth testing on low-impact tasks, with narrow workflow permissions and a human checkpoint wherever an incorrect action would be costly.