Running multi-agent workflows in n8n in production: 3 architectural lessons
Top-Explanation-4750 · reddit · 2026-11-07
A Reddit user shares months of production experience running multi-agent pipelines in n8n (research, content generation, database syncing) and the architectural bottlenecks tutorials rarely mention.
- Decouple monolithic canvases by state boundaries: single-canvas agent setups break when external APIs rate-limit or 502 — node-level retries re-run expensive LLM calls. Fix: split into single-responsibility micro-workflows via Execute Workflow nodes/webhooks, passing state explicitly through PostgreSQL/Airtable (research → ideation → validation → generation).
- Add an Evaluator/Critic gate before execution: run a cheap model (GPT-4o-mini or Gemini 1.5 Flash, temperature 0) to output binary pass/fail or confidence scores checking citations and hallucinated metrics; route records below 0.8 confidence to a manual-review queue.
- Topic overlap detection: embed new task titles and check cosine similarity against existing items in pgvector/Pinecone; above 0.82, merge or update instead of generating duplicate content.
Stack: self-hosted n8n, Gemini 1.5 Pro / Claude 3.5 Sonnet for reasoning, Flash/Mini for formatting and eval, pgvector + Airtable for triage.
More from coding & agent
- Indie dev replaces Buffer with custom HTML scheduler + AdsPower antidetect browser for multi-account posting — yihui_indie · 2026-11-07
- How devs are stopping multi-step agents from blowing up token costs — useralltime_835 · 2026-11-07
- Grok×Cursor account merge buried under 3 settings pages, users hit duplication bugs — theo · 2026-11-07
- Building a company research agent with Langflow, MCP, SEC EDGAR and Wikipedia — Aerosherm · 2026-11-07
- LlamaIndex says the world needs two types of document OCR: liteParse for speed, OCRIA-1 for accuracy — llama_index · 2026-11-07
- HDLBits: the Verilog practice site with instant simulation feedback — cneuralnetwork · 2026-11-07