Why I quit n8n for my own AI agents

It was 7 a.m. on a Tuesday morning. Coffee in hand, I checked my site before starting the day. And there it was, that little doubt: yesterday’s data wasn’t there. My n8n workflow had died overnight. One node refused to talk to an API. No clear error message. Just a red dot on a screen and a whole morning spent re-wiring blocks instead of doing my actual job.

That Tuesday morning wasn’t the first. It was just one too many.

Today, my automations run without n8n. They run on AI agents, on a small [VPS I pay €3.60/month for](https://mintavocado.com/en/hermes-agent-vps-setup/). I’m going to tell you why I switched, what n8n still gave me, and above all: in which cases you should NOT follow my example.

What n8n gave me

Let’s start by being honest, because it’s too easy to trash a tool after leaving it.

n8n taught me to think in automations. Before it, I did everything by hand: copy-pasting, repetitive checks, technical tasks I kept putting off forever. n8n showed me that you could chain steps together — a trigger, a request, a transformation, a send — and let the machine do the work.

It also made me understand the logic of connections between services. Webhooks, API keys, data formats. I learned all of that by building visual workflows, without writing a single line of code at first.

And honestly, for getting started, it’s excellent. You see your flow right in front of you, you plug blocks together, you test with one click. The first workflow that runs on its own makes you feel like a genius. It’s a legitimate gateway into the world of automation.

If I had never touched n8n, I probably would never have had the idea to build what came next.

The breaking point

So why leave? Because of one word: maintenance.

Every workflow I added was a little piece of technical debt. At first, with two or three automations, it was fine. But as soon as you have a dozen, a sneaky phenomenon kicks in: the APIs of the services you use change. One provider changes its response format, another adds authentication, a third deprecates an endpoint. And your workflow breaks.

And not in a spectacular way, either. It breaks silently, or it breaks on a Tuesday morning when you have something else to do. You open the editor, you spot the failing node, you fix it, you test, you rerun. Thirty minutes, an hour. Then the next week, another workflow. Then another.

The logic itself was frozen too. A workflow can only do what its branches say. If an unexpected case comes up — a slightly different data format, an empty response — it doesn’t adapt. It crashes. And the only intelligence capable of unblocking it was me, manually.

At some point, I did the math: I was spending more time maintaining my automations than they were saving me. An automation that costs more maintenance than it returns in time is no longer an automation. It’s a demanding pet.

What “AI agents” actually change

Here’s the heart of my decision, and it’s a difference in kind, not in degree.

A workflow automates steps. An agent reasons about a goal.

With a workflow, you tell it: “take this, transform that, send it there”. It executes the list, period. With an agent, you tell it: “write me a draft article on this topic, following these rules”. The steps are up to you. The way to get there is up to it. If an API returns something unexpected, it doesn’t crash: it adjusts, it retries, or it asks me. That’s the whole difference between a machine that breaks down and a colleague who comes back to you when there’s a problem.

Concretely, on my VPS, my agents currently handle:

• **Writing**: four articles per week, prepared as drafts, with the editorial rules respected. I review, I adjust, I approve.
• **Translation sync**: my English and Spanish articles automatically follow the French versions.
• **Automatic publishing**: approved content goes live without me touching anything.
• **Weekly SEO monitoring**: I get Search Console data and a clear summary, without ever opening a single dashboard.

What struck me most is what disappeared from my life: no more catastrophic Tuesday mornings. When an external service changes something, the agent notices, works around it, or flags it to me in one sentence. I don’t repair cables anymore. I give directions.

And maintenance? It still exists, let’s be clear. But it has shifted: instead of re-wiring visual nodes, I describe rules in natural language. And a rule written in plain language gets fixed in one sentence — not in a debugging session.

The real cost

Let’s talk money and time, because that’s where most articles stop too soon.

My current infrastructure:

• A VPS at **€3.60 per month**. That’s the price of two coffees.
• An LLM subscription at about **$15 per month**. That’s the only “modern” expense.
• A few home-made scripts and MCP connections, written over time, with the help of AI, precisely.

On the other side, my old n8n cost wasn’t a subscription. It was my time. An hour of repairs here, an evening of debugging there, the mental load of knowing a workflow could die at any moment. Even at a modest hourly rate, that time was worth far more than €20 per month.

There’s also an upfront cost — don’t lie to yourself about it. Setting up a VPS, installing your agents, writing your first scripts: it takes time, a few evenings. It’s not one-click magic like n8n. But it’s a one-time investment, where n8n maintenance was a perpetual subscription paid in hours.

So the real calculation isn’t “how much does the tool cost”. It’s “how much does the system cost, maintenance included, over twelve months”. From that angle, there’s no contest.

Who should stay on n8n

Now for the anti-dogma part, because the last thing I want is to sell you a religion.

If you have **three simple, stable automations** — a webhook that fills a Google Sheet, a Slack notification when a form is submitted, a scheduled backup — stay on n8n. Seriously. Your system runs, it’s visual, it costs a few minutes of maintenance per month. You have no problem to solve. Switching tools to be trendy is a waste.

n8n is not a bad tool. It’s probably the best entry point into automation if you don’t code. It isn’t the problem. The problem is the threshold.

The threshold is the point where the number of workflows multiplies failure surfaces faster than it multiplies your time savings. For me, that threshold was crossed somewhere around a dozen interconnected automations. For you, it will be somewhere else. The right question to ask isn’t “n8n or agents?” but “am I spending more time repairing than building?”. If the answer is no, change nothing.

And if you’re a total beginner? Start with n8n, learn the logic of connections, automate one genuinely tedious task. Then, the day maintenance starts eating you alive, you’ll know exactly why you’re switching to agents.

Conclusion

I didn’t leave n8n out of spite. I left it because my needs had outgrown what it was designed for. It taught me how to automate; my agents taught me how to delegate.

A workflow executes steps. An agent pursues a goal. When everything goes well, the difference is invisible. When the unexpected happens — and it always does — one crashes and the other adapts.

If you recognize yourself in that Tuesday morning when a workflow died during your coffee, ask yourself the question. If not, keep your n8n — it suits you just fine.

FAQ

Why quit n8n when it’s still working?

Because “working” isn’t the right unit of measurement. My workflows were working, but every external API that evolved created a repair to make. Beyond about ten automations, maintenance was consuming more time than the automations were saving. The break-even point had passed.

Do AI agents cost a lot?

Not in my case: a VPS at €3.60 per month and an LLM subscription at about $15 per month. The real cost is the initial setup time, a few evenings. In exchange, I eliminated the recurring maintenance hours n8n imposed on me every week.

Do you need to know how to code to get started with AI agents?

No, not in the classic sense. Most of my configurations are written in natural language: rules, instructions, contexts. AI helped me write the few scripts that were needed. What matters is knowing how to describe precisely what you want to get.

Is n8n a bad tool?

No, and I’m happy to say it again. For simple, stable automations — a webhook to Google Sheets, a notification, a backup — it remains excellent and visual. My departure says nothing about n8n: it says that my usage had outgrown its natural scope.

Where should I start if I want to do the same?

Start by identifying your most tedious-to-maintain automated task, not the most important one. Describe it to an AI agent as a goal rather than a sequence of steps — [the step-by-step installation guide](https://mintavocado.com/en/hermes-agent-vps-setup/) takes half an hour. Test on a single task for a few weeks, and keep n8n for everything that already runs well.