Building a Self-Healing Research Agent in n8n

Most agents repeat their mistakes. I built a self-healing research agent in n8n whose Discord users close the loop for it: every answer gets a human verdict, the verdict goes back into the same vector store the agent retrieves from, and the next run starts briefed on what worked and what didn't, with no fine-tuning anywhere, just one payload filter. I'll show the build live end-to-end, and you'll leave with an importable workflow.

1. A working pattern for self-healing agent loops: one collection as the agent's run history, its eval set, and its training signal, with the human in the loop as the writer. 2. The one-filter trick: retrieving past failures separately from past successes, so every new run is briefed on both. 3. The importable workflow plus setup README, credited and licensed, ready to run against your own Discord server.