n8n's preview Agents call existing workflows as tools for open-ended jobs like triage

A request arrives in a channel the agent has been assigned, such as Telegram or Linear ^[9]. The agent reads it and chooses its steps from the tools it has been allowed to use ^[4]. One of those tools can be a workflow the team already runs ^[5]. According to a dev.to write-up of n8n's announcement, the company is positioning agents alongside its workflow automation, not as a replacement for it ^[3].

I would defend the workflow-as-tool design at review. The agent decides whether to act, and the workflow decides how. A predictable job such as moving a completed form into a CRM stays a fixed sequence ^[14]. The agent's only choice is whether to call it. The part that varies from run to run sits at the front of the process, where someone used to read the message before anything started ^[16].

The connection also runs in the other direction. A new Message an Agent node lets a workflow invoke an agent ^[8]. That covers the mixed case among the three operating models n8n describes, where a task is partly open-ended and partly deterministic ^[11]. A fixed pipeline can hand one interpretive step, such as sorting an incoming ticket, to an agent ^[8].

Access is the constraint that shapes the rest. In a workflow, what runs is what the builder placed. An agent can reach anything on its tool list ^[4]. That list can include MCP servers and sub-agents as well as n8n's own integrations ^[5]. Each of those is a tool whose own reach the builder has to know before granting it. n8n lists per-tool credentials, approvals and scoped tools among the controls ^[7]. The write-up does not describe how an approval reaches a person or what happens when nobody answers. Its advice, which I agree with, is to start with a bounded task, carefully scoped credentials and an approval point where appropriate ^[13].

Memory and sessions carry context across interactions, and files such as CSVs and PDFs can ground the agent ^[6]. Morning summaries and email organization are among the use cases n8n's examples support ^[15]. Both depend on the agent keeping context between runs.

The demo is a Telegram assistant that turns photos of sticky notes into tasks, logs receipts and sorts email ^[10]. I have read less accurate accounts of how office work arrives.

Adoption is narrow for now. Agents are in preview on n8n Cloud for the latest stable version ^[2]. Self-hosting requires additional setup ^[12].