When n8n stops being enough: signs you need a custom agent

n8n is brilliant for the first dozen workflows. We push almost every client through it before recommending anything custom — it pays for itself within weeks and gives non-engineers visibility into how automation works.

But there's a point where the fit breaks. Here are the five signs we look for, in order of urgency.

First — when the same workflow has more than three or four conditional branches, and any of them touches AI. n8n handles conditional logic fine, but debugging an LLM response inside a Switch node when something goes wrong is genuinely painful. Custom code with proper logging wins here every time.

Second — when a workflow needs to run reliably at high concurrency. n8n's execution model is fine for hundreds of runs per day, awkward for thousands per hour, and broken for tens of thousands. If you're consistently queuing jobs, you've outgrown it.

Third — when your data shape is unpredictable. AI agents that browse the web, read documents, or extract structured data from unstructured input often produce variable schemas. Validation logic in n8n quickly becomes the workflow.

Fourth — when you need real evaluation harnesses. Once an agent is making decisions that touch revenue, you need an eval suite that compares output against gold standards. n8n has no native fixture story; custom code does.

Fifth — when the workflow is core enough that it deserves real engineering rigour: version control on logic (not just JSON exports), code review on changes, CI/CD on deploys, observability on every step. n8n can be pushed there but it fights you.