What 72 agents building software actually looks like
My software factory runs on loops nested inside loops. I wanted to see what one big, successful task actually looks like from the inside, so I charted every loop.
One orchestrator agent directed 71 subagents through about 15 active hours of work and shipped 22 pull requests of significant new functionality. That’s 72 agents on one task.
The clock draws every loop in it: me steering the orchestrator, the orchestrator’s work episodes, each workstream’s execute, judge, fix and re-judge cycle, and each agent’s own loop of tool calls.
That task was a big effort, and it wouldn’t have been possible without engineering the loops.
Open the clock and find the red curls. Each one is a judge rejecting work and sending it back. Tell me what you see.
