It is becoming f*cking dangerous A 15-year-old schoolboy from China found instructions on how to launch 300 Claude Code agents so that while he’s at school they would trade, and he connected them to a verification layer so that none of them could screw him over. 200 agents collect and write data simultaneously Every claim comes with a link to a snippet of the source A separate, cheaper model checks whether the source actually confirms the fact Failed check → the agent redoes it. Still not confirmed after 3 attempts → goes into a separate list for a human Everything that passes immediately drops into an Obsidian vault as notes The main trick is role separation. One model does the work, another only decides “true or false.” Verifying a fact against a source is much simpler and cheaper than finding it in the first place A team of analysts would bill for weeks for this. He did it in a single run The same Second Brain + swarm scheme that keeps circulating