From Monolith to Multi-Agent: Rebuilding Nate on n8n

Quickly resolving simple helpdesk tickets is a great use case for n8n. But when a single AI agent handles every ticket type, the system gets harder to maintain, and you lose the ability to specialize, evaluate, and scale capabilities independently.

This talk covers how n8n's support engineering team rebuilt Nate, the AI assistant embedded in our customer support workflow, from a single generalist into a team of specialist agents — built and tested on n8n itself.

We'll walk through the Nate 2.0 architecture: the intent classifier and its three-level, 10-category taxonomy powering the router, the classifier-to-specialist pipeline, and per-category playbooks — markdown runbooks fetched at runtime so agent behavior changes are easy to test.

We'll show how peer-to-peer agents-as-tools handle tickets spanning multiple intents, cover inter-agent handoffs, and look at safeguards around high-risk actions like subscription cancellations: identity verification before any destructive tool call, and idempotency guards against repeat actions.

Finally, we'll show how the end-to-end test suite runs as an n8n workflow, with test cases in n8n data tables, mocked tool results so evals never touch real infrastructure, and an LLM judge grading replies against per-ticket rubrics.