Lessons from running multi-agent workflows in n8n in production
After months of running multi-agent pipelines in n8n — research, content generation, database syncing — one builder shares the architectural lessons tutorials skip. Connecting LLM nodes, vector stores and tool-calling sub-nodes on a single canvas looks fine in testing until real workloads hit. The core lesson: split by job, not by tool, give each agent its own memory, and add a human approval step before anything writes to a database or sends externally.