60 legal tech newsletters hit my inbox every day.

I leverage agents to only show me the 3 most relevant articles from all of them.

People ask how a CEO has time to read this much, post here a couple times a week and still do my day job.

The answer is a pipeline that reads for me and tees up what matters.

For the past year that pipeline has been a dedicated Gmail account subscribed to every legal tech newsletter I could find, plus a chain of n8n workflows.

Extract the URLs, crawl the articles, have DeepSeek score each one against a rubric I wrote, log everything to a Google Sheet, email me anything that scores 70 or higher.

It worked. It also needed babysitting. n8n upgrades. Topping up the DeepSeek balance. A broken node every few weeks/months.

The bigger problem is that it never got smarter. The rubric I wrote in month one is the rubric it ran in month twelve.

I tried replacing it with OpenClaw, Hermes, Copilot, and cron jobs in Claude Code. Each one added a different kind of friction and I kept falling back to n8n.

Today I rebuilt the whole thing on GrokBot with a single agent. No Google Sheet. No n8n. No Firecrawl. No separate scoring model.

The agent reads the inbox, scores against my rubric, and sends the digest. Screenshots below.

On its first test slice it cleared seven articles above 70 and dropped two that didn't belong, which is exactly what the filter is supposed to do.

I can now tell it "that scoring was too generous" in plain English and it adjusts. The old pipeline couldn't learn. This one can.

Each bot runs in its own VM you can log into with a browser. Bots can talk to each other.

I'd only ever used Grok for testing before this. Early signs are strong.