Building n8n’s AI Assistant: Wrong Turns and Lessons Learned

What began as a scrappy two-engineer effort to "build an n8n super agent" turned into an assistant that explained errors and answered questions — and has since evolved into a system that can understand an n8n instance, build and modify workflows, coordinate multi-step tasks, and verify its own work.

This talk tells the engineering story behind that evolution. We'll walk through the major versions of n8n's AI Assistant — from a simple conversational interface, through the first workflow builder, to multi-agent experiments and today's more autonomous architecture — examining what worked, what broke as usage and ambition grew, and why the architecture had to change each time.

We'll cover real challenges: giving models enough context without overwhelming them, generating valid workflows across hundreds of nodes, coordinating agents without unnecessary complexity, recovering long-running operations, handling permissions and human approval, and evaluating a system whose output is inherently non-deterministic.

This is a candid account of the trade-offs, failed approaches, and unexpected lessons behind turning an AI prototype into a real, launched product.