Let Them Run

I found out about Hermes from a YouTuber with a few hundred thousand views. To my credit, I traced it back to a 30-year-old PhD thesis within an hour. But that is not the edge I want. The edge is reading the paper before anyone makes the video. So I built a swarm of agents that does the slog for me, runs all night, and turns a corpus into novel theses I can be first to test.

When founders in regulated industries try to point an LLM at the data that actually matters, they hit the Compliance Wall. This post walks through Layer 1 of getting past it: a four-primitive architecture (recently named the Reasoner-Executor-Synthesizer pattern in the agent literature) running end-to-end on synthetic European medical records, hosted in Amsterdam, with an LLM driving and never seeing a single real name.

If you're already running an AI Second Brain on markdown plus grep, your search layer is quietly underperforming. After auditing 50+ memory repos, I picked the one that ships. Here's what changes when you swap grep for hybrid search.

I pointed 12 AI agents at Accelerando and extracted 230 business ideas in seven minutes. But the real goldmine wasn't the ideas - it was the mental models. I extracted Stross's post-scarcity reasoning patterns, tested them against my savings decisions, and compared them to traditional frameworks. The synthesis was better than either alone.

I asked my AI second brain to extract knowledge from a year of meeting transcripts. It mined 133 conversations into 40+ contact profiles and 9,500 lines of structured intelligence. Then it built me a replacement for the transcription tool I was paying for. Overnight. While I slept.

A client candidly admitted he couldn't show AI-generated work to his customers. That confession gave me the key to unlock a different skeptic, one who'd been resisting for weeks. Here's how my AI second brain turned a casual insight into a research-backed argument in under ten minutes.

How AI swarms turn design budgets into exploration, not deliverables.

I let 86 AI agents write a book about swarm intelligence, then fed the book back to the swarm as its own upgrade plan, and it worked.

I didn't write a single paragraph of this book. I described an outcome, went to sleep, and woke up to 431,000 words of working material distilled into a 54,000-word tactical field guide.

Everyone's running Claude Code in Cursor. I was too. So I built an AI workforce manager where I describe the outcome, deploy swarms overnight, and only get pulled back in when they're stuck.

The execution bottleneck that has kept great ideas trapped in busywork is dissolving and what's left is pure creative potential.