I was intrigued by an AMD issue on GitHub where Stella Laurenzo shared how the AMD team runs a fleet of 50 or so agents:
https://lnkd.in/drueskiV. I wanted to get there myself, and I think I managed to. I can now run a fleet of coding agents while I'm away, burning through every token across my three Claude subscriptions. I can run data science experiments overnight in loops: analyze misclassifications, form a hypothesis, implement, run, analyze again. Agents pull tasks from a queue, add new tasks back to it, organize dependencies, and block tasks when they need my input. This article walks through my approach step by step.
https://github.com/sermakarevich/claude/tree/main/usage_patt...
