n8n's New AI Assistant Lets You Build Automation Workflows by Chatting — The 'Vibe Coding for Workflows' Era Begins

TL;DR

n8n's new AI Assistant (July 9) replaces the old AI Workflow Builder with a Claude Opus 4.8-powered chat agent that builds, edits, tests, and troubleshoots workflows from plain English. Three days later, n8n shipped a companion AI Agent (July 12) that runs autonomously alongside your workflows. I tested both against real automation scenarios. The chat-to-build experience is strong for simple integrations, decent for mid-complexity workflows, but it stumbles on n8n-specific patterns and edge cases. Zapier and Make should be paying attention.

What It Is

On July 9, n8n shipped something that changes how you interact with the platform at a fundamental level. The AI Assistant lives in the left sidebar of n8n Cloud (version 2.29.9 and above). You pick a project, describe what you want to automate, and it builds the workflow. It can also edit, test, and troubleshoot existing ones.

Under the hood, it runs on Claude Opus 4.8. n8n's launch FAQ confirms: "on n8n Cloud it currently uses opus 4.8 for most of the stuff." That's a serious reasoning engine, not a template-matcher.

The older AI Workflow Builder is gone. The FAQ is direct: "The AI Assistant can do everything the workflow builder can do and more, and with better results. As such it replaces the AI Workflow Builder." No transition period.

On the safety side, it respects project permissions, never sends your API keys to the AI, and asks for confirmation before publishing or deleting. Sensible guardrails for a preview feature.

Testing It: Simple Stuff

I started with something basic: a workflow that checks Gmail for emails from a specific sender, extracts attachments, saves them to Google Drive, and pings me on Slack with a summary.

The Assistant proposed a plan in about 15 seconds. Gmail trigger, sender filter, attachment extraction, Drive upload, Slack notification. It asked for confirmation, then built the thing.

Clean. Node configurations were correct. It even added an error-handling branch I hadn't asked for. First test execution worked. This would've taken me maybe 20 minutes manually. I had it running in under two. This is where the Assistant earns its keep: integrations where the mapping between what you want and what n8n nodes exist is straightforward.

Testing It: Complex Agents

Then I threw something harder at it: a lead qualification agent that receives form submissions, enriches company data via Clearbit, scores leads based on industry and company size, routes high-scoring leads to Slack and HubSpot, and sends low-scoring leads a different email sequence.

The node mapping was right again. Clearbit enrichment, routing logic, HubSpot deal creation: all present. But the scoring logic went into a single Function node with JavaScript. It worked, but it wasn't idiomatic n8n. Want to tweak scoring thresholds later? You're digging through a Function node instead of using Switch nodes where the logic is visible.

When I asked it to refactor into Switch nodes, it did, but the result was verbose. Eight Switch nodes where three would do. The Assistant is good at getting things working; it's less good at getting them elegant.

Then I asked it to add exponential backoff retry for the Clearbit API call. It added a Wait node and a Loop node with linear backoff. I had to explain what exponential backoff means before it corrected itself. If you know n8n well, you'll catch these things. If you don't, you might ship a workflow with hidden performance problems.

Where It Stumbles

The Assistant is in preview, and the rough edges are real:

It sometimes asks clarifying questions on things that don't need clarifying. "Which Google Drive folder?" is fine. "Would you like the Slack message to contain text?" is absurd. The clarification loop burns credits without adding value.

It doesn't know n8n idioms. Code node versus Function node. "Execute workflow" versus "call sub-workflow." HTTP Request node versus native integration. The Assistant picks one, but not necessarily the right one for the situation. If you're new to n8n, you won't know to second-guess it.

Credits go fast. Long conversations, big workflows, repeated iterations: it all adds up. n8n says they're working on letting users buy credits without upgrading plans, but for now, heavy users on lower tiers will hit the wall.

Self-hosted setup is manual and not smooth. If you run n8n on your own infra, expect a separate setup guide and some fiddling.

The Plot Twist: AI Agent on July 12

Three days after the Assistant, n8n shipped a second AI feature: an AI Agent that runs natively alongside workflows. Different beast. The Assistant helps you build; the Agent is the workflow (or part of it). It can make decisions during execution, reason through variable scenarios, and interact with the same 400+ integrations.

It lives in the same interface as traditional workflows. You can mix structured logic with agentic decision-making in a single automation: deterministic data processing, then hand off to the agent for a subjective judgment (is this support ticket actually urgent?), then back to deterministic routing.

This two-pronged launch tells you where n8n is heading. They're betting the line between builder and runtime blurs, and the same AI that constructs workflows will eventually operate inside them.

Versus Zapier and Make

Zapier Copilot builds Zaps from natural language too. It's fine for simple app-to-app connections, and 7,000+ apps gives it breadth. But it feels like a guided template selector. It maps intent to existing templates. n8n's Assistant generates from scratch, which gives it more room for custom logic.

Make is investing in MCP gateway infrastructure, letting AI agents discover and call tools through a standardised protocol. Different philosophy: Make builds plumbing for AI-to-automation connections. n8n puts the AI inside the builder.

Practical difference: with n8n, you chat. With Zapier, you describe and select. With Make, you configure and connect. n8n's approach is the most conversational, and therefore the most accessible to people who know what they want but not how to build it.

The Bigger Picture

When the tool that builds your automations gets smart enough to build them for you, the skill shifts from "knowing how to configure nodes" to "knowing what's worth automating and how to describe it clearly." Lower barrier, higher ceiling.

But it also means the platforms that absorb the most AI complexity win. If your team has to think about model routing and credit optimisation just to automate invoice processing, the tool hasn't done its job. Platforms like Stacker take a different approach: absorb the AI complexity so business teams build internal tools with permissions, portals, and workflows without ever touching a model selector or a Function node. That's where the whole space needs to go.

Verdict

Who it's for: Teams already using n8n who want to accelerate workflow creation. Technical builders who know enough n8n to spot when the Assistant makes a suboptimal call. Not yet for someone expecting AI to handle everything perfectly on the first go.

Pros: Builds working workflows from plain English fast. Claude Opus 4.8 gives it real reasoning power. Respects permissions and never touches credentials. The companion AI Agent extends the vision from building to running. Self-hostable.

Cons: Stumbles on n8n idioms and elegant workflow design. Credit consumption is opaque. Self-hosted setup is manual. Preview means behaviour may change.

The takeaway: n8n's AI Assistant is the most aggressive chat-to-automation move from a major platform. It's not finished, but it's not vapour. The real test isn't whether it can build a workflow. It can. The test is whether n8n can make it reliable enough that you trust it with production automations without babysitting every node it generates. There's still work to do.