The Workflow Engineer

Most engineers underrate workflow automation because they confuse it with low-code marketing demos. They picture a marketing manager dragging a line between a Facebook Lead Ads node and a Mailchimp node, clapping when…

Pick the wrong workflow tool and you'll pay for it for years. Most teams pick by what they've used before, then justify it retroactively. There's a better way.

Most n8n workflows I'm asked to debug have the same shape: an organic accretion of IF nodes, Set nodes, and one heroic Code node that's load-bearing.

n8n's "getting started" guides are fine for kicking the tires. They are not fine for running anything you care about. Here's what production actually looks like.

The Code node is the most powerful primitive in n8n and the one most teams use wrong. Either they avoid it entirely and end up with 14-node Set chains that snake across the canvas like spaghetti, or they shove…

There's a kind of Code node I see in almost every workflow I'm asked to debug: 200 lines, no comments, doing six unrelated things. It is, technically, working.

Webhooks are the part of n8n where toy examples are most misleading. The webhook node in a tutorial is a 5-minute setup. The webhook node in a production system is the difference between data integrity and pages at 3am.

Most workflow bugs aren't logic bugs. They're shape bugs — a field is null when you expected a string, or an array when you expected an object. The fix isn't more null checks; it's a different way of thinking about data.

Two-way sync looks like one of the easiest workflow problems and is in fact one of the hardest. Most teams get the demo working in an afternoon and spend the next year fixing edge cases.

Most teams discover their API bills the way I discover mine: an unexpectedly large invoice. Then they spend an afternoon panicking, find one obvious cut, and stop.

Most AI-in-workflow tutorials skip the parts that make AI production-grade. They show you the happy path: prompt goes in, smart-sounding text comes out, ship it.

If you're parsing AI output with regex, you're holding a system together with tape. If you're parsing it with JSON.parse and a try/catch, you're closer.

Most n8n workflows handle errors the way most code does: a try/catch wrapped around the obvious place, an alert that fires when something is already broken, and nothing else. That's not error handling.

Workflow developers talk about testing the way back-end developers talked about it in 2008: nice in theory, never the priority. The result is the same — most workflows get tested in production by their users, often…

n8n performs well at small scale and at large scale. It is in the middle — the messy "we have real users now" range — where most teams hit walls they didn't expect.

Most self-hosted n8n setups I'm asked to review look the same: a docker-compose.yml from a tutorial, a Postgres with default settings, and someone's personal email for alerts. They work until they don't.

Workflow security has a credibility problem. "Use strong passwords" is the level of advice most teams get. Meanwhile, the real failure modes — credential sprawl, code-node escape, unsanitized webhook input — go…

There’s a moment in every workflow project where you have to choose: build one more node into the monolith, or pull a chunk into a sub-workflow. Most teams keep adding nodes.

AI phone agents are easy to demo and brutal to run in production. Voice has a latency budget the chatbot world doesn't. Get it wrong and the caller hangs up.

Customer support is where AI in workflow gets real fastest. The volumes are high, the cost of bad answers is high, and the upside of getting it right is enormous.

Most "AI content pipelines" are demos. They generate plausible text, post it, and wait for the cringe to surface. The pipelines that actually work look different. They don't move text from prompt to platform.

Sales teams change CRMs. They always do. The question isn't whether you'll migrate; it's whether your automation survives the migration. Most don't. Mine do.

Monitoring a workflow stack is a different problem from monitoring a service. The unit of work is a workflow execution, not a request. The failure modes are different.

n8n ships hundreds of nodes. Most of them you'll never use. The handful that earn permanent residency in my workflows have things in common — and the ones I avoid do too.

Every n8n developer hits the same wall at some point: "why is this expression returning undefined?" The answer is almost always the same — they're thinking about the data shape wrong, and n8n's expression language…

If you've landed here from a search result, you're probably one of three readers: the engineer who just inherited a workflow system, the founder building one for the first time, or the consultant doing the work for…