AI agent or n8n? My simple rule for choosing
Mike Codeur
Claude Code killed no-code. AI agents killed traditional automation. We have been repeating it for months and, broadly, it is true.
I build my SaaS products with Claude Code. I run OpenClaw and Hermes on tasks I was wiring inside Make two years ago. The underlying shift is real: no-code experts are coming back to code, and visual workflows are being replaced by plain language.
Except I kept n8n. And not out of nostalgia.
Why no-code and automation won in the first place
Before AI, development was the bottleneck. Slow and expensive. An MVP cost tens of thousands and took months.
Two families of tools exploded at the same time to work around that:
- Product no-code: Bubble, Webflow, Airtable, Glide. Build your app without writing code.
- Automation: Zapier, Make, n8n, Pipedream. Connect your tools without writing scripts.
They did not win because they were technically better. They won because development was too slow.
That matters, because if the reason they won disappears, so does their advantage.
What changed on the product side
Claude Code is not autocomplete. It is an agent working inside your repository: it reads the project, edits files, runs commands, reads errors and fixes itself.
In practice, developers get back the one thing no-code genuinely had: speed. A landing page with an email capture form is ninety seconds of prompting instead of thirty minutes of drag and drop.
Except at the end you get what Bubble never gave you:
- a real Git repository, with history
- tests
- an architecture you choose
- hosting you control
- the ability to leave without rewriting everything
No-code used to say "I am faster than you". Today the developer answers "so am I, and I end up with a real product".
What changed on the automation side
Make and Zapier rest on the same principle: you wire your flow node by node, by hand. Twelve nodes for logic you could have described in one sentence.
With an agent, you write that sentence. "Sort my inbox: archive spam, reply to welcome messages, assign partnership requests." The agent handles routing, decision, action and iteration on its own.
The real difference is not syntax, it is comprehension. A Make scenario applies patterns you wired in advance. An agent understands context, including the cases you did not anticipate.
For anything custom and moderately intelligent, the agent wins. Clearly.
The pricing trap when it works
Almost all these tools share one thing: usage-based billing. Per operation for Make, per workload unit for Bubble, per record times per seat for Airtable.
You start at a few euros a month. Then your product works, volume grows, and the invoice follows the same curve. You can end up paying hundreds a month for logic you could have hosted yourself for a fraction of that.
A custom project on Vercel and Supabase stays broadly flat as volume grows. The curve does not run away.
No-code punishes you at the exact moment your product starts working. That is the opposite of what you want.
Where the real line sits
The debate is no longer no-code versus code. It is elsewhere:
stochastic versus deterministic
A language model is stochastic. Same prompt, same context, the result can vary from one run to the next. You can constrain it with rules and examples: you reduce the variance, you do not remove it.
For the vast majority of what I ask an agent, that variance is a strength. It is what I am paying for: the ability to handle a case I did not anticipate.
For the rest, it is a risk I do not take.
What I never hand to an agent
Some flows are ones where a different result is not intelligent adaptation, it is an incident:
| Flow | Why not an agent |
|---|---|
| Stripe webhooks, payment capture | A mishandled payment is an accounting incident, not a variation |
| CRM synchronisation | A duplicated or lost customer record pollutes the entire sales chain |
| Compliance and GDPR logs | You must be able to prove processing was identical every time |
| Any recurring critical process | If the result changes, you can no longer debug it |
These flows need three things an agent does not guarantee: a reproducible result, an observable execution, and predictable retries.
That is the definition of a workflow engine. n8n, but also Temporal or Airflow depending on scale.
This is also why n8n survives where Make and Zapier retreat: n8n was never no-code for non-developers. It is a workflow engine for developers, self-hostable, versionable, with usable logs. It is not the same product category.
My 2026 stack
I no longer choose between them. I stack them.
Intelligence layer, stochastic. Claude Code, Cursor, the OpenClaw and Hermes agents. For everything that requires understanding context and making a judgement call.
Product layer. Next.js, Supabase, Stripe, Vercel. The custom repository, the one that scales without the invoice exploding.
Critical flow layer, deterministic. Self-hosted n8n. For everything that must produce exactly the same thing on every run.
Airtable, Notion and Make remain side tools. Useful occasionally, never a foundation.
The rule
If you remember one thing:
- Must the flow produce an identical result on every run? Workflow engine, so n8n.
- Must the flow understand context and make a judgement call? AI agent.
- Is it the product itself? Code, written with Claude Code.
The mistake in 2026 is not choosing n8n over an agent. It is believing you have to choose.
I go through each point with on-screen demos in the video: AI agent or n8n, my simple rule