A couple of weeks ago I shared a small workflow: it turns YouTube videos I find interesting into structured research docs, so the good ideas don't evaporate the moment the video ends. New tech, architecture talks, build logs. Transcribe it, break it down, and file it where my agents and I can both find it later, so we can ideate on it and figure out how it fits the stack we are already building. That workflow was the goal. I built it because I wanted it, and it does its job. But finishing it left me holding something I didn't have before: a research library that grows every time I digest something good. And that asset spawned the next idea. Now that I've got this dataset, what can I automate around it? So I built a LangChain blog writer agent on top of it. Every 2 weeks it reads that same research library, surfaces three things worth writing about, and hands me the shortlist. I pick one. It drafts the post. I approve it. Then it publishes to my site on its own. I keep exactly two decisions: which idea, and whether the draft is good enough to ship. The agent does the rest. That's how I approach AI. You don't architect the grand system up front. You build one honest thing because it's worth having, and the finished thing hands you raw material you didn't have yesterday. One good idea becomes the dataset for the next one. It cascades. Now here's the part I left out, because the cascade goes back further than the YouTube workflow. None of this was the plan. It started when I decided to actually learn LangGraph and LangSmith. Not for a project. Just to understand how agents get built, and how you trace what they do once they are running. I made toy graphs. I sat and read the traces. I broke things on purpose to see what the tooling would show me. That curiosity was the first domino. Learning the framework is what got me building agents at all. Building agents is what made me want a research pipeline to feed them. That pipeline is what made the blog writer obvious.

Using LangChain agents to diagnose and repair n8n workflow errors autonomously. I built this (and 8 other Langchain agents) to wow the hiring manager at Langchain in my interview with them this morning I run a lot of workflows in n8n. When one fails, a global error handler catches it and hands it to a LangGraph agent. The agent diagnoses the failure and drafts a fix with Anthropic, then passes that fix to a different model from OpenAI to review and approve before anything is written. That second step is the whole point, and the design is model agnostic. The proposer and the reviewer are just two seats, and you can drop any model into either one. Right now Claude drafts the fix and an OpenAI model approves it, so no single model marks its own homework. The fix only goes live after that independent, cross-vendor review. Then the agent applies it through the n8n API and re-runs the workflow to confirm it actually worked. In the demo I break a workflow on purpose and let the agent repair it end to end. It triaged the error, proposed the fix, got it approved, wrote it back, and the workflow ran clean. The whole thing cost under 6 cents, and every step is traceable in LangSmith. I triggered this run by hand so it was easy to follow, but the agent is set to run automatically the moment a new error lands. It is still in testing, and I am keeping it that way on purpose, but the autonomous loop is the point. That last part matters more than the fix. When agents act on their own inside your systems, you have to see exactly what they did. Observability is not a nice-to-have, it is the architecture. . #AIAgents #n8n #LangGraph #AIOps #Automation