How we built olddogs.ai with lots of agents but zero slop