On September 25, 2026, n8n shipped something it probably should have shipped a year ago: native AI Agents that live next to your workflows, not inside them. If you have used n8n before, you know the AI Agent node — the thing you wire up with a chat trigger, some memory, and a handful of tools, then spend an afternoon debugging because one connection came loose. That node still works. Nothing broke. But the new Agents feature is a different animal entirely, and it changes what n8n is for.
What actually changed
Until yesterday, building an agent in n8n meant assembling the plumbing yourself. Chat trigger, memory node, AI Agent node, tool connections, error handling, session management — all of it hand-wired on the canvas. It worked, but it was the equivalent of building your own PC from parts. You picked every component and connected every cable. If you knew what you were doing, you got a great machine. If you did not, you got a very expensive pile of components that beeped at you.
n8n Agents are the pre-built version. You describe what the agent should do in plain language, pick a model, attach the tools and workflows it can use, and the agent figures out the rest. Memory, sessions, channels, versioning, and approvals are baked in. You spend your time on what the agent should do, not on how to wire it up.
The key design decision is that agents and workflows are peers. Your agent can call your existing workflows as tools — so a customer support agent can trigger your refund workflow, or a sales agent can run your lead enrichment pipeline. And when a workflow needs an agent for one of its steps, the new Message an Agent node calls it from inside the workflow. Both directions work. The agent does not replace workflows; it sits next to them and cooperates.
Why this matters for real work
The reason n8n built this now is that models have gotten good enough at figuring out multi-step plans on their own. Two years ago, you had to tell the model exactly what to do, step by step, because it would reliably mess up anything open-ended. Today, you can give a current model a goal and the right tools, and it is genuinely good at working out the sequence.
That matters most for jobs where the input is different every time. Someone in Slack asks why a customer's usage dropped last month. A support email comes in with a billing question that touches three systems. A GitHub issue needs triage across two repositories and a Notion board. You cannot lay out those processes in advance because the next step depends on the answer to the last one. With agents, you describe the goal, give it access to what it needs, and let it figure out the path.
The Slack integration is the part that caught my attention. You can talk to your agent directly in Slack, the same agent that also runs on a schedule and can be called from any workflow. One agent, three entry points, same behavior. That is a meaningful upgrade for teams that live in Slack and do not want to open another tool to check on something.
What this is not
This is not n8n pivoting to become a chatbot platform. The agents are still automation-first. They run on schedules, they get called from workflows, they have structured inputs and outputs. The Slack interface is a convenience layer, not the product. If you want a pure conversational AI, there are better tools. If you want an AI that can actually do things in your systems — send emails, update databases, trigger webhooks, run your existing n8n workflows — this is the most integrated option I have seen.
It is also not a replacement for the AI Agent node. If you have existing agents built with that node, they keep working. The new Agents feature is a higher-level abstraction that includes memory and session management out of the box, so you do not have to wire those up yourself anymore.
The pricing question
n8n Agents are available on all plans, including self-hosted. If you run n8n on your own server, you get agents for free — you just pay for the model API calls. On n8n Cloud, agents count toward your execution quota like any other workflow. There is no separate "agent tax" on top of what you already pay.
The real cost is the model. Running an agent that thinks through multi-step plans burns more tokens than a simple workflow with a single LLM call. If you are using GPT-4 or Claude for your agent, expect each interaction to cost somewhere between $0.02 and $0.15 depending on complexity. For a high-volume support agent handling hundreds of conversations a day, that adds up. For a sales assistant that triages ten leads a morning, it is negligible.
What to do with this
If you are already on n8n, the first thing to try is wrapping one of your existing workflows in an agent. Pick something you have built that has a few decision points — lead routing, support triage, content scheduling — and give the agent access to that workflow as a tool. Describe the goal in plain language and see how it handles the edge cases you currently handle with if/else nodes.
If you are not on n8n yet, this is a good reason to look at it. The self-hosted option means you own your data, the execution costs are predictable, and now you get agents without paying a platform premium for the privilege. The learning curve is real — n8n is not as simple as Zapier — but the payoff at scale is significantly better.
The automation market has been quietly splitting into two camps: tools that are great at connecting apps in fixed sequences, and tools that are trying to let AI figure out the sequence for you. n8n just planted a flag in both camps simultaneously. Workflows for the predictable stuff, agents for the messy stuff, and a single platform that runs both. That is a strong position to be in heading into 2027.
Source: Introducing n8n Agents — n8n Blog, September 25, 2026