n8n has introduced Agents, a new type of building block that sits alongside its existing workflow canvas rather than replacing it. Instead of laying out fixed steps, you describe in plain language what an agent should do, attach a model, give it channels and triggers (Slack, Telegram, Linear, Discord, or a schedule), and let it work out the sequence itself.
The core change for automation builders is that workflows and agents can now call each other directly. An agent can use any existing n8n workflow as a tool, so a support agent might have three underlying workflows: pull account context, add a note to an account, or page on-call. The agent decides when to call each one; what the workflow itself does stays fixed exactly as built. In the other direction, a new Message an Agent node lets any workflow hand a step off to a published agent and get its answer back, using the same instructions, tools and memory the agent uses everywhere else.
Controls are built in per tool: any tool can be marked sensitive so the agent pauses for a human approve/reject before using it, and each tool runs on its own attached credential rather than giving the agent broad access to the underlying system. Agents also get drafts and published versions (edit and test a draft while the team keeps using the published one), and full session logs showing every step, tool call and input/output.
Other components of an agent include skills (reusable instructions and reference files, shareable across agents), sub-agents (one agent calling another), and knowledge (uploaded csv, pdf, markdown or txt files, plus vector stores on n8n Cloud). The existing AI Agent node is unchanged and continues to work for anyone already using it.
Agents are available now on n8n Cloud for everyone on the latest stable version, with self-hosted support requiring extra setup and Enterprise availability described as coming soon. On cost, one agent turn counts as one workflow execution, tool calls to workflows and sub-agents don't count separately, and agent executions share the account's existing workflow execution quota; building an agent through n8n's Assistant uses AI credits like any Assistant conversation. n8n describes the feature as still in preview and is asking users to test before publishing and keep approvals on sensitive actions.